From f311d92d5e723f677281d5cd72a815b2ca46f2a4 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 15:08:50 +0200
Subject: [PATCH 01/38] docs: enrich model registry JSDoc comments and document
 rnexecutorchJsi

---
 .../react-native-executorch/src/models.ts     | 281 +++++++++++-------
 .../src/native/bridge.ts                      |   9 +
 packages/react-native-executorch/src/utils.ts |   4 +-
 3 files changed, 178 insertions(+), 116 deletions(-)

diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index b4db68689c..72ea6530f2 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -1333,9 +1333,9 @@ export const models = {
   classification: {
     /**
      * EfficientNetV2-S image classification model pre-trained on ImageNet-1k
-     * (1000 categories, see {@link IMAGENET1K_LABELS}).
-     * Compact and efficient architecture providing high accuracy for
-     * general-purpose image classification.
+     * (1000 categories, see {@link IMAGENET1K_LABELS}). Compact and efficient
+     * architecture providing high accuracy for general-purpose image
+     * classification.
      */
     EFFICIENTNET_V2_S: {
       ...EFFICIENTNET_V2_S_XNNPACK_INT8,
@@ -1361,8 +1361,8 @@ export const models = {
       COREML_FP16: STYLE_TRANSFER_CANDY_COREML_FP16,
     },
     /**
-     * Fast neural style transfer model applying a classic tile mosaic
-     * artistic pattern.
+     * Fast neural style transfer model applying a classic tile mosaic artistic
+     * pattern.
      */
     MOSAIC: {
       ...STYLE_TRANSFER_MOSAIC_XNNPACK_INT8,
@@ -1371,8 +1371,8 @@ export const models = {
       COREML_FP16: STYLE_TRANSFER_MOSAIC_COREML_FP16,
     },
     /**
-     * Fast neural style transfer model applying a painterly "Rain Princess"
-     * oil painting aesthetic.
+     * Fast neural style transfer model applying a painterly "Rain Princess" oil
+     * painting aesthetic.
      */
     RAIN_PRINCESS: {
       ...STYLE_TRANSFER_RAIN_PRINCESS_XNNPACK_INT8,
@@ -1393,14 +1393,14 @@ export const models = {
   },
 
   /**
-   * Semantic segmentation models that classify each pixel into target object
-   * or background classes.
+   * Semantic segmentation models that classify each pixel into target object or
+   * background classes.
    */
   semanticSegmentation: {
     /**
      * Lightweight portrait selfie segmentation model for real-time person vs
-     * background separation. Categorizes pixels into `background` and
-     * `person`. Ideal for background blur and replacement effects.
+     * background separation. Categorizes pixels into `background` and `person`.
+     * Ideal for background blur and replacement effects.
      */
     SELFIE_SEGMENTATION: {
       ...SELFIE_SEGMENTATION_XNNPACK_FP32,
@@ -1428,8 +1428,8 @@ export const models = {
       COREML_FP16: LRASPP_MOBILENET_V3_LARGE_COREML_FP16,
     },
     /**
-     * DeepLabV3 semantic segmentation model with ResNet-50 backbone
-     * (21 classes, see {@link PASCAL_VOC_LABELS}). High-accuracy segmentation
+     * DeepLabV3 semantic segmentation model with ResNet-50 backbone (21
+     * classes, see {@link PASCAL_VOC_LABELS}). High-accuracy segmentation
      * utilizing atrous spatial pyramid pooling.
      */
     DEEPLAB_V3_RESNET50: {
@@ -1439,8 +1439,8 @@ export const models = {
       COREML_FP16: DEEPLAB_V3_RESNET50_COREML_FP16,
     },
     /**
-     * DeepLabV3 semantic segmentation model with ResNet-101 backbone
-     * (21 classes, see {@link PASCAL_VOC_LABELS}). High-capacity backbone for
+     * DeepLabV3 semantic segmentation model with ResNet-101 backbone (21
+     * classes, see {@link PASCAL_VOC_LABELS}). High-capacity backbone for
      * maximum segmentation detail and boundary accuracy.
      */
     DEEPLAB_V3_RESNET101: {
@@ -1450,8 +1450,8 @@ export const models = {
       COREML_FP16: DEEPLAB_V3_RESNET101_COREML_FP16,
     },
     /**
-     * DeepLabV3 semantic segmentation model with MobileNetV3-Large backbone
-     * (21 classes, see {@link PASCAL_VOC_LABELS}). Combines DeepLabV3 feature
+     * DeepLabV3 semantic segmentation model with MobileNetV3-Large backbone (21
+     * classes, see {@link PASCAL_VOC_LABELS}). Combines DeepLabV3 feature
      * extraction quality with a lightweight mobile backbone.
      */
     DEEPLAB_V3_MOBILENET_V3_LARGE: {
@@ -1507,16 +1507,16 @@ export const models = {
       COREML_FP16: RFDETR_NANO_DETECTOR_COREML_FP16,
     },
     /**
-     * Ultralytics YOLO26 real-time object detection models trained on COCO
-     * (80 classes, see {@link COCO_CLASSES_YOLO}). Available across multiple
-     * scale sizes (NANO, SMALL, MEDIUM, LARGE, XLARGE) and resolutions
-     * (384x384, 512x512, 640x640).
+     * Ultralytics YOLO26 real-time object detection models trained on COCO (80
+     * classes, see {@link COCO_CLASSES_YOLO}). Available across multiple scale
+     * sizes (NANO, SMALL, MEDIUM, LARGE, XLARGE) and resolutions (384x384,
+     * 512x512, 640x640).
      */
     YOLO26: {
       ...YOLO26_NANO_384_XNNPACK_FP32,
       /**
-       * Nano scale YOLO26 object detection model. High speed,
-       * ultra low latency.
+       * Nano scale YOLO26 object detection model. High speed, ultra low
+       * latency.
        */
       NANO: {
         ...YOLO26_NANO_384_XNNPACK_FP32,
@@ -1534,8 +1534,8 @@ export const models = {
         },
       },
       /**
-       * Small scale YOLO26 object detection model. Balanced
-       * latency and accuracy.
+       * Small scale YOLO26 object detection model. Balanced latency and
+       * accuracy.
        */
       SMALL: {
         ...YOLO26_SMALL_384_XNNPACK_FP32,
@@ -1553,8 +1553,8 @@ export const models = {
         },
       },
       /**
-       * Medium scale YOLO26 object detection model. Higher
-       * precision for complex scenes.
+       * Medium scale YOLO26 object detection model. Higher precision for
+       * complex scenes.
        */
       MEDIUM: {
         ...YOLO26_MEDIUM_384_XNNPACK_FP32,
@@ -1572,8 +1572,7 @@ export const models = {
         },
       },
       /**
-       * Large scale YOLO26 object detection model. High accuracy
-       * model variant.
+       * Large scale YOLO26 object detection model. High accuracy model variant.
        */
       LARGE: {
         ...YOLO26_LARGE_384_XNNPACK_FP32,
@@ -1591,8 +1590,8 @@ export const models = {
         },
       },
       /**
-       * Extra Large scale YOLO26 object detection model.
-       * Maximum detection performance.
+       * Extra Large scale YOLO26 object detection model. Maximum detection
+       * performance.
        */
       XLARGE: {
         ...YOLO26_XLARGE_384_XNNPACK_FP32,
@@ -1618,17 +1617,18 @@ export const models = {
    */
   keypointDetection: {
     /**
-     * MediaPipe BlazeFace lightweight face detection and 6-point facial landmark
-     * locator (eyes, nose, mouth, ears, see {@link BLAZEFACE_LANDMARKS}).
+     * MediaPipe BlazeFace lightweight face detection and 6-point facial
+     * landmark locator (eyes, nose, mouth, ears, see
+     * {@link BLAZEFACE_LANDMARKS}).
      */
     BLAZEFACE: {
       ...BLAZEFACE_XNNPACK_FP32,
       XNNPACK_FP32: BLAZEFACE_XNNPACK_FP32,
     },
     /**
-     * YOLO26 human pose estimation model predicting 17 COCO body keypoints
-     * (see {@link COCO_LANDMARKS}). Available across 384x384, 512x512, and
-     * 640x640 resolutions.
+     * YOLO26 human pose estimation model predicting 17 COCO body keypoints (see
+     * {@link COCO_LANDMARKS}). Available across 384x384, 512x512, and 640x640
+     * resolutions.
      */
     YOLO26_POSE: {
       ...YOLO26_POSE_384_XNNPACK_FP32,
@@ -1649,8 +1649,8 @@ export const models = {
   },
 
   /**
-   * Instance segmentation models predicting both bounding boxes and fine-grained
-   * pixel masks per object instance.
+   * Instance segmentation models predicting both bounding boxes and
+   * fine-grained pixel masks per object instance.
    */
   instanceSegmentation: {
     /**
@@ -1660,8 +1660,7 @@ export const models = {
      */
     FASTSAM: {
       /**
-       * FastSAM Small - lightweight instance segmenter for
-       * mobile.
+       * FastSAM Small - lightweight instance segmenter for mobile.
        */
       S: {
         ...FASTSAM_S_XNNPACK_FP32,
@@ -1678,9 +1677,9 @@ export const models = {
       },
     },
     /**
-     * RF-DETR (Roboflow Detection Transformer) Nano instance segmentation
-     * model predicting COCO class masks and bounding boxes
-     * (see {@link COCO_CLASSES}).
+     * RF-DETR (Roboflow Detection Transformer) Nano instance segmentation model
+     * predicting COCO class masks and bounding boxes (see
+     * {@link COCO_CLASSES}).
      */
     RFDETR_NANO: {
       ...RFDETR_NANO_SEG_COREML_FP16,
@@ -1696,8 +1695,8 @@ export const models = {
     YOLO26: {
       ...YOLO26_NANO_SEG_384_XNNPACK_FP32,
       /**
-       * Nano scale YOLO26 instance segmentation model. High
-       * speed.
+       * Nano scale YOLO26 instance segmentation model. High speed, ultra low
+       * latency mask generation.
        */
       NANO: {
         ...YOLO26_NANO_SEG_384_XNNPACK_FP32,
@@ -1705,28 +1704,40 @@ export const models = {
         SIZE_512: { XNNPACK_FP32: YOLO26_NANO_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_NANO_SEG_640_XNNPACK_FP32 },
       },
-      /** Small scale YOLO26 instance segmentation model. */
+      /**
+       * Small scale YOLO26 instance segmentation model. Balanced latency and
+       * mask accuracy.
+       */
       SMALL: {
         ...YOLO26_SMALL_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_SMALL_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_SMALL_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_SMALL_SEG_640_XNNPACK_FP32 },
       },
-      /** Medium scale YOLO26 instance segmentation model. */
+      /**
+       * Medium scale YOLO26 instance segmentation model. Higher mask boundary
+       * precision for complex multi-object scenes.
+       */
       MEDIUM: {
         ...YOLO26_MEDIUM_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_640_XNNPACK_FP32 },
       },
-      /** Large scale YOLO26 instance segmentation model. */
+      /**
+       * Large scale YOLO26 instance segmentation model. High accuracy instance
+       * segmentation variant for demanding visual pipelines.
+       */
       LARGE: {
         ...YOLO26_LARGE_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_LARGE_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_LARGE_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_LARGE_SEG_640_XNNPACK_FP32 },
       },
-      /** Extra Large scale YOLO26 instance segmentation model. */
+      /**
+       * Extra Large scale YOLO26 instance segmentation model. Maximum instance
+       * segmentation and mask delineation performance.
+       */
       XLARGE: {
         ...YOLO26_XLARGE_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_XLARGE_SEG_384_XNNPACK_FP32 },
@@ -1763,8 +1774,8 @@ export const models = {
      */
     WHISPER: {
       /**
-       * Multilingual Whisper Tiny model. Supporting 99+
-       * languages. High speed speech recognition.
+       * Multilingual Whisper Tiny model. Supporting 99+ languages. High speed
+       * speech recognition.
        */
       TINY: {
         ...WHISPER_TINY_XNNPACK_FP32,
@@ -1774,8 +1785,8 @@ export const models = {
         MLX_INT8: WHISPER_TINY_MLX_INT8,
       },
       /**
-       * Multilingual Whisper Base model. Higher accuracy across
-       * supported languages.
+       * Multilingual Whisper Base model. Higher accuracy across supported
+       * languages.
        */
       BASE: {
         ...WHISPER_BASE_XNNPACK_FP32,
@@ -1785,8 +1796,8 @@ export const models = {
         MLX_INT8: WHISPER_BASE_MLX_INT8,
       },
       /**
-       * Multilingual Whisper Small model. Best accuracy for
-       * complex multi-language audio.
+       * Multilingual Whisper Small model. Best accuracy for complex
+       * multi-language audio.
        */
       SMALL: {
         ...WHISPER_SMALL_XNNPACK_FP32,
@@ -1797,8 +1808,7 @@ export const models = {
       /** English-only optimized Whisper models (`TINY`, `BASE`, `SMALL`). */
       EN: {
         /**
-         * English-only Whisper Tiny model. Fast and compact for
-         * English STT.
+         * English-only Whisper Tiny model. Fast and compact for English STT.
          */
         TINY: {
           ...WHISPER_TINY_EN_XNNPACK_FP32,
@@ -1808,8 +1818,8 @@ export const models = {
           MLX_INT8: WHISPER_TINY_EN_MLX_INT8,
         },
         /**
-         * English-only Whisper Base model. High accuracy
-         * English speech recognition.
+         * English-only Whisper Base model. High accuracy English speech
+         * recognition.
          */
         BASE: {
           ...WHISPER_BASE_EN_XNNPACK_FP32,
@@ -1819,8 +1829,8 @@ export const models = {
           MLX_INT8: WHISPER_BASE_EN_MLX_INT8,
         },
         /**
-         * English-only Whisper Small model. Superior accuracy
-         * for English transcription.
+         * English-only Whisper Small model. Superior accuracy for English
+         * transcription.
          */
         SMALL: {
           ...WHISPER_SMALL_EN_XNNPACK_FP32,
@@ -1841,13 +1851,16 @@ export const models = {
   },
 
   /**
-   * Generative Large Language Models (LLMs) for instruction following,
-   * chat, text generation, and reasoning.
+   * Generative Large Language Models (LLMs) for instruction following, chat,
+   * text generation, and reasoning.
    */
   llm: {
     /**
-     * Liquid AI LFM 2.5 1.2B general-purpose text model. Excellent for complex
-     * on-device reasoning, instruction following, and fast multi-turn chat.
+     * Liquid AI LFM 2.5 1.2B general-purpose hybrid language model. Built on
+     * the Liquid Foundation Model architecture for low memory bandwidth usage,
+     * and high-throughput token generation. Delivers strong general-purpose
+     * reasoning, instruction following, and fast multi-turn conversational chat
+     * on mobile devices.
      */
     LFM2_5_1_2B: {
       ...LFM2_5_1_2B_XNNPACK_8DA4W,
@@ -1856,8 +1869,10 @@ export const models = {
       MLX_INT4: LFM2_5_1_2B_MLX_INT4,
     },
     /**
-     * Liquid AI LFM 2.5 350M ultra-compact text model. Best for low-latency text
-     * completion, quick responses, and resource-constrained devices.
+     * Liquid AI LFM 2.5 350M ultra-compact hybrid language model. Optimized for
+     * minimal memory footprint and sub-second first-token response times. Ideal
+     * for lightweight text completion, fast intent classification, query
+     * routing, and low-latency chat on resource-constrained edge hardware.
      */
     LFM2_5_350M: {
       ...LFM2_5_350M_XNNPACK_8DA4W,
@@ -1866,8 +1881,10 @@ export const models = {
       MLX_INT4: LFM2_5_350M_MLX_INT4,
     },
     /**
-     * Liquid AI LFM 2.5 450M vision-language model. Optimized for real-time
-     * visual QA, image description, and low-latency multimodal chat.
+     * Liquid AI LFM 2.5 VL 450M lightweight multimodal vision-language model.
+     * Combines Liquid hybrid language modeling with visual token embeddings for
+     * real-time on-device visual question answering (VQA), image description,
+     * UI element inspection, and low-latency multimodal conversational agents.
      */
     LFM2_5_VL_450M: {
       ...LFM2_5_VL_450M_XNNPACK_8DA4W,
@@ -1876,8 +1893,10 @@ export const models = {
       VULKAN_8DA4W: LFM2_5_VL_450M_VULKAN_8DA4W,
     },
     /**
-     * Liquid AI LFM 2.5 1.6B vision-language model. Higher quality visual
-     * understanding, detailed image analysis, and complex multimodal tasks.
+     * Liquid AI LFM 2.5 VL 1.6B high-capacity vision-language model. Provides
+     * fine-grained visual scene understanding, document/chart interpretation,
+     * detailed image captioning, and multi-turn visual dialogue with higher
+     * precision and reasoning fidelity than the 450M variant.
      */
     LFM2_5_VL_1_6B: {
       ...LFM2_5_VL_1_6B_XNNPACK_8DA4W,
@@ -1885,8 +1904,10 @@ export const models = {
       VULKAN_8DA4W: LFM2_5_VL_1_6B_VULKAN_8DA4W,
     },
     /**
-     * Bielik v3 1.5B Polish & English language model. Fine-tuned specifically for
-     * native Polish fluency, grammar, and bilingual translation.
+     * Bielik v3 1.5B bilingual Polish & English language model, developed by
+     * SpeakLeash. Fine-tuned on curated Polish corpora and instruction datasets
+     * for native Polish cultural nuance, grammar accuracy, idioms, and
+     * high-fidelity bidirectional Polish-English translation.
      */
     BIELIK_V3_1_5B: {
       ...BIELIK_V3_1_5B_XNNPACK_8DA4W,
@@ -1894,8 +1915,11 @@ export const models = {
       XNNPACK_FP16: BIELIK_V3_1_5B_XNNPACK_FP16,
     },
     /**
-     * Meta Llama 3.2 1B multilingual text model. Ideal for lightweight mobile
-     * chat, summary generation, and multilingual prompt processing.
+     * Meta Llama 3.2 1B lightweight instruction-tuned multilingual model.
+     * Features Grouped-Query Attention (GQA) and SpinQuant quantization for
+     * compact memory utilization and high throughput. Well suited for on-device
+     * text summarization, prompt rewriting, and lightweight conversational
+     * assistance.
      */
     LLAMA3_2_1B: {
       ...LLAMA3_2_1B_SPINQUANT,
@@ -1903,8 +1927,10 @@ export const models = {
       XNNPACK_BF16: LLAMA3_2_1B_BF16,
     },
     /**
-     * Meta Llama 3.2 3B multilingual text model. Strong instruction following,
-     * detailed content creation, and high-precision text reasoning.
+     * Meta Llama 3.2 3B instruction-tuned multilingual language model. Delivers
+     * strong instruction adherence, multi-turn reasoning, and high-quality
+     * content creation across 8+ core languages while maintaining a compact
+     * on-device memory profile.
      */
     LLAMA3_2_3B: {
       ...LLAMA3_2_3B_SPINQUANT,
@@ -1912,8 +1938,9 @@ export const models = {
       XNNPACK_BF16: LLAMA3_2_3B_BF16,
     },
     /**
-     * Hugging Face SmolLM2 135M sub-parameter model. Best for micro-footprint
-     * background tasks, simple text tagging, and instant autocomplete.
+     * Hugging Face SmolLM2 135M ultra-compact language model. Engineered for
+     * micro-memory footprints, instant token generation, text classification,
+     * and background processing on low-power devices.
      */
     SMOLLM2_135M: {
       ...SMOLLM2_135M_8DA4W,
@@ -1921,8 +1948,10 @@ export const models = {
       XNNPACK_BF16: SMOLLM2_135M_BF16,
     },
     /**
-     * Hugging Face SmolLM2 360M compact model. Balanced speed and intelligence
-     * for lightweight conversational assistants.
+     * Hugging Face SmolLM2 360M compact instruction-tuned model. Provides a
+     * practical balance between fast mobile generation speed and conversational
+     * coherence, ideal for lightweight on-device assistants, text
+     * simplification, and structured data extraction.
      */
     SMOLLM2_360M: {
       ...SMOLLM2_360M_8DA4W,
@@ -1930,8 +1959,10 @@ export const models = {
       XNNPACK_BF16: SMOLLM2_360M_BF16,
     },
     /**
-     * Hugging Face SmolLM2 1.7B language model. Powerful general-purpose text
-     * generation, creative writing, and general knowledge Q&A.
+     * Hugging Face SmolLM2 1.7B language model trained on curated educational,
+     * synthetic, and web data. Delivers competitive reasoning, creative text
+     * generation, and general knowledge Q&A performance approaching larger
+     * 2B-3B models while maintaining fast on-device inference.
      */
     SMOLLM2_1_7B: {
       ...SMOLLM2_1_7B_8DA4W,
@@ -1939,8 +1970,10 @@ export const models = {
       XNNPACK_BF16: SMOLLM2_1_7B_BF16,
     },
     /**
-     * Hammer 2.1 0.5B function-calling model. Specialized for lightweight agentic
-     * tool calling, JSON extraction, and structured output parsing.
+     * Hammer 2.1 0.5B specialized function-calling model. Fine-tuned
+     * specifically for agentic tool use, structured JSON extraction, and
+     * single/multi-tool invocation with ultra-low latency for real-time mobile
+     * tool calling flows.
      */
     HAMMER2_1_0_5B: {
       ...HAMMER2_1_0_5B_XNNPACK_8DA4W,
@@ -1948,8 +1981,9 @@ export const models = {
       XNNPACK_BF16: HAMMER2_1_0_5B_XNNPACK_BF16,
     },
     /**
-     * Hammer 2.1 1.5B function-calling model. Optimized for multi-tool agentic
-     * workflows, API function calling, and structured JSON schemas.
+     * Hammer 2.1 1.5B function-calling language model. Optimized for multi-tool
+     * agentic workflows, API parameter schema validation, and structured JSON
+     * output generation on edge devices.
      */
     HAMMER2_1_1_5B: {
       ...HAMMER2_1_1_5B_XNNPACK_8DA4W,
@@ -1957,8 +1991,10 @@ export const models = {
       XNNPACK_BF16: HAMMER2_1_1_5B_XNNPACK_BF16,
     },
     /**
-     * Hammer 2.1 3B function-calling model. High-capacity agentic reasoning,
-     * complex multi-step tool execution, and robust schema compliance.
+     * Hammer 2.1 3B high-capacity function-calling model. Provides top-tier
+     * tool selection precision, multi-turn tool calling, error recovery, and
+     * strict compliance with complex TypeScript/JSON schema specifications in
+     * autonomous mobile agent pipelines.
      */
     HAMMER2_1_3B: {
       ...HAMMER2_1_3B_XNNPACK_8DA4W,
@@ -1966,8 +2002,10 @@ export const models = {
       XNNPACK_BF16: HAMMER2_1_3B_XNNPACK_BF16,
     },
     /**
-     * Microsoft Phi-4 Mini 3.8B reasoning model. Exceptional for math problem
-     * solving, logical reasoning, code synthesis, and analytical tasks.
+     * Microsoft Phi-4 Mini 3.8B high-density reasoning model. Trained on
+     * synthetic textbook-grade datasets for state-of-the-art on-device STEM
+     * problem solving, complex mathematical reasoning, multi-step code
+     * synthesis, and structured analytical tasks.
      */
     PHI4_MINI: {
       ...PHI4_MINI_XNNPACK_8DA4W,
@@ -1975,8 +2013,10 @@ export const models = {
       XNNPACK_BF16: PHI4_MINI_XNNPACK_BF16,
     },
     /**
-     * Alibaba Qwen 2.5 0.5B multilingual model. Extremely efficient for fast
-     * multi-language translation and basic conversational chat.
+     * Alibaba Qwen 2.5 0.5B ultra-lightweight multilingual model. Trained on
+     * 18T tokens supporting 29+ languages; optimized for near-instant response
+     * times, basic instruction following, multilingual translation, and
+     * lightweight conversational assistants on mobile devices.
      */
     QWEN2_5_0_5B: {
       ...QWEN2_5_0_5B_XNNPACK_8DA4W,
@@ -1984,8 +2024,10 @@ export const models = {
       XNNPACK_BF16: QWEN2_5_0_5B_XNNPACK_BF16,
     },
     /**
-     * Alibaba Qwen 2.5 1.5B multilingual model. Great for balanced multilingual
-     * chat, text summarization, and cross-lingual understanding.
+     * Alibaba Qwen 2.5 1.5B multilingual instruction model. Combines broad
+     * multilingual comprehension across 29+ languages with strong coding and
+     * math capabilities, well suited for interactive chat, summarization, and
+     * cross-lingual translation.
      */
     QWEN2_5_1_5B: {
       ...QWEN2_5_1_5B_XNNPACK_8DA4W,
@@ -1993,8 +2035,9 @@ export const models = {
       XNNPACK_BF16: QWEN2_5_1_5B_XNNPACK_BF16,
     },
     /**
-     * Alibaba Qwen 2.5 3B multilingual model. High capability across 29+
-     * languages for complex translation, long-form writing, and Q&A.
+     * Alibaba Qwen 2.5 3B high-capability multilingual model. Delivers strong
+     * reasoning, coding, mathematics, and multilingual fluency across 29+
+     * languages for in-depth text generation and complex multi-turn dialogue.
      */
     QWEN2_5_3B: {
       ...QWEN2_5_3B_XNNPACK_8DA4W,
@@ -2002,8 +2045,10 @@ export const models = {
       XNNPACK_BF16: QWEN2_5_3B_XNNPACK_BF16,
     },
     /**
-     * Alibaba Qwen 3 0.6B next-gen text model. Low-latency multilingual model
-     * for fast turn-taking and concise response generation.
+     * Alibaba Qwen 3 0.6B next-generation compact language model. Features
+     * updated architectural optimizations for reduced latency, enhanced
+     * multilingual token representation, and efficient conversational
+     * turn-taking on mobile devices.
      */
     QWEN3_0_6B: {
       ...QWEN3_0_6B_XNNPACK_8DA4W,
@@ -2011,8 +2056,9 @@ export const models = {
       XNNPACK_BF16: QWEN3_0_6B_XNNPACK_BF16,
     },
     /**
-     * Alibaba Qwen 3 1.7B next-gen text model. Versatile multilingual assistant
-     * for high-quality instruction following and knowledge retrieval.
+     * Alibaba Qwen 3 1.7B next-generation multilingual language model. Balances
+     * high reasoning capability, general knowledge retrieval, coding
+     * proficiency, and conversational fluidity across multiple languages.
      */
     QWEN3_1_7B: {
       ...QWEN3_1_7B_XNNPACK_8DA4W,
@@ -2020,8 +2066,10 @@ export const models = {
       XNNPACK_BF16: QWEN3_1_7B_XNNPACK_BF16,
     },
     /**
-     * Alibaba Qwen 3 4B high-capacity text model. Top-tier multilingual
-     * reasoning, technical content generation, and multi-turn dialogue.
+     * Alibaba Qwen 3 4B high-capacity generative model. Delivers advanced
+     * multi-step reasoning, comprehensive world knowledge, complex coding
+     * capabilities, and top-tier multilingual performance for demanding
+     * on-device AI applications.
      */
     QWEN3_4B: {
       ...QWEN3_4B_XNNPACK_8DA4W,
@@ -2029,8 +2077,10 @@ export const models = {
       XNNPACK_BF16: QWEN3_4B_XNNPACK_BF16,
     },
     /**
-     * Google Gemma 4 E2B generative text model. Built on Google's Gemini tech
-     * for high-fidelity instruction following and mobile assistance.
+     * Google Gemma 4 E2B generative language model. Built on Google's Gemini
+     * research and architecture innovations, offering high-fidelity instruction
+     * following, creative text generation, and reasoning efficiency optimized
+     * for mobile deployment.
      */
     GEMMA4_E2B: {
       ...GEMMA4_E2B_XNNPACK_8DA4W,
@@ -2062,8 +2112,8 @@ export const models = {
       XNNPACK_FP32: ALL_MPNET_BASE_V2_EMBEDDINGS,
     },
     /**
-     * 384-dimensional sentence transformer fine-tuned specifically for
-     * semantic QA matching using cosine similarity.
+     * 384-dimensional sentence transformer fine-tuned specifically for semantic
+     * QA matching using cosine similarity.
      */
     MULTI_QA_MINILM_L6_COS_V1: {
       ...MULTI_QA_MINILM_L6_COS_V1_EMBEDDINGS,
@@ -2095,9 +2145,9 @@ export const models = {
       MLX_INT8: DISTILUSE_BASE_MULTILINGUAL_CASED_V2_MLX_INT8,
     },
     /**
-     * CLIP text encoder (ViT-B/32) mapping text queries into a
-     * 512-dimensional joint text-image embedding space. Used in combination
-     * with `imageEmbeddings.CLIP_VIT_BASE_PATCH32` for zero-shot text-to-image
+     * CLIP text encoder (ViT-B/32) mapping text queries into a 512-dimensional
+     * joint text-image embedding space. Used in combination with
+     * `imageEmbeddings.CLIP_VIT_BASE_PATCH32` for zero-shot text-to-image
      * search.
      */
     CLIP_VIT_BASE_PATCH32_TEXT: {
@@ -2165,8 +2215,8 @@ export const models = {
   textToImage: {
     /**
      * Ultra-fast SDXS (Stable Diffusion eXtreme Speed) 512x512 text-to-image
-     * generation model based on DreamShaper. Generates high-quality images
-     * from text prompts in real time.
+     * generation model based on DreamShaper. Generates high-quality images from
+     * text prompts in real time.
      */
     SDXS_512_DREAMSHAPER: {
       ...SDXS_512_DREAMSHAPER_XNNPACK_FP32,
@@ -2176,11 +2226,14 @@ export const models = {
   },
 
   /**
-   * Text-to-Speech (TTS) models that synthesize audio waveforms from input text.
+   * Text-to-Speech (TTS) models that synthesize audio waveforms from input
+   * text.
    */
   textToSpeech: {
     /**
-     * Supertonic 3 multilingual flow-matching Text-to-Speech model.
+     * Supertonic 3 multilingual flow-matching Text-to-Speech model. Delivers
+     * natural, highly expressive speech synthesis with configurable speaker
+     * voice presets (see {@link SUPERTONIC_DEFAULT_VOICE_NAMES}).
      */
     SUPERTONIC: {
       ...SUPERTONIC_3_XNNPACK_FP32,
diff --git a/packages/react-native-executorch/src/native/bridge.ts b/packages/react-native-executorch/src/native/bridge.ts
index c3d5f6bb73..3d17ab8fc0 100644
--- a/packages/react-native-executorch/src/native/bridge.ts
+++ b/packages/react-native-executorch/src/native/bridge.ts
@@ -7,6 +7,15 @@ if (!globalObj.__rnexecutorch_jsi__) {
   if (NativeRnExecutorch) NativeRnExecutorch.install();
 }
 
+/**
+ * Direct reference to the native C++ JSI host object (`__rnexecutorch_jsi__`).
+ *
+ * Provides low-level, synchronous bindings for ExecuTorch core runtime
+ * operations (model loading, tensor allocation and manipulation, registered
+ * backends inspection, emulator detection) and native extension namespaces
+ * (`cv`, `llm`, `math`, `nlp`, `speech`).
+ * @internal
+ */
 export const rnexecutorchJsi = globalObj.__rnexecutorch_jsi__;
 
 if (!rnexecutorchJsi) {
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index d7ad54f65c..278b223be1 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -30,8 +30,8 @@ export function getRegisteredBackends(): string[] {
  * returned local path if you also intend to run the model.
  * @category Utils
  * @param source The remote HTTP URL or local path to the `.pte` model file.
- * @returns A promise resolving to an object containing the model source,
- * method signature metadata, and per-method backend usage.
+ * @returns A promise resolving to an object containing the model source, method
+ * signature metadata, and per-method backend usage.
  */
 export async function inspectModel(source: string): Promise<{
   source: string;

From 23d08eac3da002f2e49b7cc9b2f0827fad4e023b Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 15:27:02 +0200
Subject: [PATCH 02/38] refactor(hooks): unify JSDoc documentation, link
 imperative APIs and types, and use undefined for absent state

---
 .agents/skills/add-task-pipeline/SKILL.md     | 17 +++-----
 .../src/hooks/useClassifier.ts                | 16 +++++---
 .../src/hooks/useImageEmbedder.ts             | 16 +++++---
 .../src/hooks/useInstanceSegmenter.ts         | 16 +++++---
 .../src/hooks/useKeypointDetector.ts          | 16 +++++---
 .../src/hooks/useLLMChatSession.ts            | 26 ++++++------
 .../src/hooks/useModel.ts                     | 28 ++++++-------
 .../src/hooks/useObjectDetector.ts            | 18 ++++----
 .../hooks/useOpticalCharacterRecognizer.ts    | 19 +++++----
 .../src/hooks/usePrivacyFilter.ts             | 17 ++++----
 .../src/hooks/useResourceDownload.ts          | 18 ++++----
 .../src/hooks/useSemanticSegmenter.ts         | 16 +++++---
 .../src/hooks/useSpeechToText.ts              | 18 ++++----
 .../src/hooks/useStyleTransfer.ts             | 14 ++++---
 .../src/hooks/useTextEmbedder.ts              | 17 ++++----
 .../src/hooks/useTextToImage.ts               | 19 +++++----
 .../src/hooks/useTextToSpeech.ts              | 41 +++++++++++--------
 .../src/hooks/useTokenizer.ts                 | 14 ++++---
 .../src/hooks/useVoiceActivityDetector.ts     | 15 ++++---
 packages/react-native-executorch/src/utils.ts |  4 +-
 20 files changed, 212 insertions(+), 153 deletions(-)

diff --git a/.agents/skills/add-task-pipeline/SKILL.md b/.agents/skills/add-task-pipeline/SKILL.md
index 84574c50c8..d0d9d306a6 100644
--- a/.agents/skills/add-task-pipeline/SKILL.md
+++ b/.agents/skills/add-task-pipeline/SKILL.md
@@ -190,28 +190,21 @@ Wrap the task pipeline in a custom React Hook using the core hooks `useResourceD
 
 ```typescript
 import { useModel } from './useModel';
-import { useResourceDownload } from './useResourceDownload';
+import { useResourceDownload, type ResourceOptions } from './useResourceDownload';
 import { createMyTask, type MyTaskModel } from '../extensions/<domain>/tasks/<task>';
 
-export function useMyTask(config: MyTaskModel, options?: { preventLoad?: boolean }) {
+export function useMyTask(config: MyTaskModel, options?: ResourceOptions) {
   // 1. Resolve remote or local asset model path and download progress
-  const { localPath, downloadProgress, downloadError } = useResourceDownload(
-    config.modelPath,
-    options?.preventLoad
-  );
+  const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
 
   // 2. Instantiate and compile the task pipeline (with automatic lifecycle cleanup)
-  const { model, error } = useModel(
-    createMyTask,
-    localPath ? { ...config, modelPath: localPath } : null,
-    [localPath]
-  );
+  const { model, error } = useModel(createMyTask, resource);
 
   return {
     isReady: !!model,
     error: downloadError || error,
     downloadProgress,
-    localPath,
+    resource,
     runTask: model?.runTask,
     runTaskWorklet: model?.runTaskWorklet,
   };
diff --git a/packages/react-native-executorch/src/hooks/useClassifier.ts b/packages/react-native-executorch/src/hooks/useClassifier.ts
index 9612539b88..795e010589 100644
--- a/packages/react-native-executorch/src/hooks/useClassifier.ts
+++ b/packages/react-native-executorch/src/hooks/useClassifier.ts
@@ -5,20 +5,24 @@ import { createClassifier, type ClassifierModel } from '../extensions/cv/tasks/c
 /**
  * React hook to load and run an image classification model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createClassifier}.
  * @category Hooks
  * @typeParam L The type representing the classification labels.
- * @param config The image classification model configuration.
+ * @param config The image classification model configuration. See {@link
+ * ClassifierModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
- * progress, and classification functions.
+ * progress, labels, and classification functions.
+ * @see {@link createClassifier}
  */
 export function useClassifier<L>(config: ClassifierModel<L>, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createClassifier<L>, resource ?? null);
+  const { model, error } = useModel(createClassifier<L>, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useImageEmbedder.ts b/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
index 5a6eec7c17..5d91c0bfa2 100644
--- a/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
+++ b/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
@@ -6,21 +6,25 @@ import {
 } from '../extensions/cv/tasks/imageEmbedding';
 
 /**
- * React hook to load and run an image embedder model.
+ * React hook to load and run an image embedding model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createImageEmbedder}.
  * @category Hooks
- * @param config The image embedder model configuration.
+ * @param config The image embedder model configuration. See {@link
+ * ImageEmbedderModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and embedding functions.
+ * @see {@link createImageEmbedder}
  */
 export function useImageEmbedder(config: ImageEmbedderModel, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createImageEmbedder, resource ?? null);
+  const { model, error } = useModel(createImageEmbedder, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
index e986709a5d..b43b6eba25 100644
--- a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
@@ -9,24 +9,28 @@ import {
 /**
  * React hook to load and run an instance segmentation model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createInstanceSegmenter}.
  * @category Hooks
  * @typeParam F The bounding box format.
  * @typeParam L The class labels type.
- * @param config The instance segmentation model configuration.
+ * @param config The instance segmentation model configuration. See {@link
+ * InstanceSegmenterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
- * progress, and segmentation functions.
+ * progress, labels, and instance segmentation functions.
+ * @see {@link createInstanceSegmenter}
  */
 export function useInstanceSegmenter<F extends BoxFormat, L>(
   config: InstanceSegmenterModel<F, L>,
   options?: ResourceOptions
 ) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createInstanceSegmenter<F, L>, resource ?? null);
+  const { model, error } = useModel(createInstanceSegmenter<F, L>, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
index e14b10c53a..01bba9c1c1 100644
--- a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
@@ -9,24 +9,28 @@ import {
 /**
  * React hook to load and run a keypoint detection model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createKeypointDetector}.
  * @category Hooks
  * @typeParam F The bounding box format.
  * @typeParam L The landmark labels type.
- * @param config The keypoint detection model configuration.
+ * @param config The keypoint detection model configuration. See {@link
+ * KeypointDetectorModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
- * progress, and keypoint detection functions.
+ * progress, landmarks, and keypoint detection functions.
+ * @see {@link createKeypointDetector}
  */
 export function useKeypointDetector<F extends BoxFormat, L extends PropertyKey>(
   config: KeypointDetectorModel<F, L>,
   options?: ResourceOptions
 ) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createKeypointDetector<F, L>, resource ?? null);
+  const { model, error } = useModel(createKeypointDetector<F, L>, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useLLMChatSession.ts b/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
index 32b537a456..c33a6a6c76 100644
--- a/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
+++ b/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
@@ -7,26 +7,28 @@ import {
 } from '../extensions/llm/tasks/llmChatSession';
 
 /**
- * React hook to load and run an LLM chat session model.
+ * React hook to load and run an LLM chat session.
  *
- * This hook manages downloading (if they are remote URLs) and loading the `.pte` model
- * file, `tokenizer.json`, and `tokenizer_config.json`, tracking download progress and errors,
- * and cleaning up native memory when the component unmounts or configuration changes.
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets and tokenizer files, tracking download progress and load errors,
+ * and releasing native memory when the component unmounts or the configuration
+ * changes.
+ *
+ * For imperative usage, see {@link createLLMChatSession}.
  * @category Hooks
- * @param config The LLM model configuration.
- * @param options Chat session options and load/caching options. See {@link ResourceOptions}.
- * @returns An object containing the session's loading state, error, download progress,
- * and chat functions.
+ * @param config The LLM model configuration. See {@link LLMModel}.
+ * @param options Chat session options and load/caching options. See {@link
+ * ResourceOptions}.
+ * @returns An object containing the session's loading state, error, download
+ * progress, and chat functions.
+ * @see {@link createLLMChatSession}
  */
 export function useLLMChatSession(
   config: LLMModel,
   options?: LLMChatSessionOptions & ResourceOptions
 ) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model: session, error } = useModel(
-    (res) => createLLMChatSession(res, options),
-    resource ?? null
-  );
+  const { model: session, error } = useModel((res) => createLLMChatSession(res, options), resource);
 
   return {
     isReady: !!session,
diff --git a/packages/react-native-executorch/src/hooks/useModel.ts b/packages/react-native-executorch/src/hooks/useModel.ts
index aa1f022b7e..98748f7807 100644
--- a/packages/react-native-executorch/src/hooks/useModel.ts
+++ b/packages/react-native-executorch/src/hooks/useModel.ts
@@ -1,7 +1,7 @@
 import { useEffect, useMemo, useState } from 'react';
 
 /**
- * React hook to instantiate and compile a model pipeline with automatic
+ * React hook to instantiate and manage a model or task instance with automatic
  * lifetime management.
  *
  * It manages the lifecycle of a model or task runner instance. When the
@@ -11,10 +11,10 @@ import { useEffect, useMemo, useState } from 'react';
  * @typeParam TConfig The configuration type passed to the model creator.
  * @typeParam TModel The type of the compiled model instance containing a
  * `dispose` method.
- * @param createModel An asynchronous factory function to instantiate the
- * model/task.
- * @param config The configuration to pass to `createModel`, or `null` if the
- * model shouldn't be loaded yet. It is tracked by value, so the model is
+ * @param createModel An asynchronous factory function to instantiate the model
+ * or task runner.
+ * @param config The configuration to pass to `createModel`, or `undefined` if
+ * the model shouldn't be loaded yet. It is tracked by value, so the model is
  * re-created whenever the config's contents change and passing an inline object
  * is safe.
  * @returns An object containing the loaded model instance and any instantiation
@@ -22,26 +22,26 @@ import { useEffect, useMemo, useState } from 'react';
  */
 export function useModel<TConfig, TModel extends { dispose: () => void }>(
   createModel: (config: TConfig) => Promise<TModel>,
-  config: TConfig | null
+  config: TConfig | undefined
 ) {
-  const [model, setModel] = useState<TModel | null>(null);
-  const [error, setError] = useState<Error | null>(null);
+  const [model, setModel] = useState<TModel | undefined>();
+  const [error, setError] = useState<Error | undefined>();
 
   // Configs are plain JSON data, so serializing is a sound structural identity
   // and keeps an inline `config` object from rebuilding the model every render.
   const configKey = useMemo(() => JSON.stringify(config), [config]);
 
   useEffect(() => {
-    if (!config) {
-      setModel(null);
-      setError(null);
+    if (config === undefined) {
+      setModel(undefined);
+      setError(undefined);
       return;
     }
 
     let isMounted = true;
-    let instance: TModel | null = null;
-    setModel(null);
-    setError(null);
+    let instance: TModel | undefined;
+    setModel(undefined);
+    setError(undefined);
 
     createModel(config)
       .then((m) => {
diff --git a/packages/react-native-executorch/src/hooks/useObjectDetector.ts b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
index 6a724840cf..e66109628a 100644
--- a/packages/react-native-executorch/src/hooks/useObjectDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
@@ -9,24 +9,28 @@ import {
 /**
  * React hook to load and run an object detection model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createObjectDetector}.
  * @category Hooks
- * @typeParam L The type representing the object class labels.
  * @typeParam F The bounding box format.
- * @param config The object detection model configuration.
+ * @typeParam L The type representing the object class labels.
+ * @param config The object detection model configuration. See {@link
+ * ObjectDetectorModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
- * progress, and object detection functions.
+ * progress, labels, and object detection functions.
+ * @see {@link createObjectDetector}
  */
 export function useObjectDetector<F extends BoxFormat, L>(
   config: ObjectDetectorModel<F, L>,
   options?: ResourceOptions
 ) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createObjectDetector<F, L>, resource ?? null);
+  const { model, error } = useModel(createObjectDetector<F, L>, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useOpticalCharacterRecognizer.ts b/packages/react-native-executorch/src/hooks/useOpticalCharacterRecognizer.ts
index 3b2ee02a3f..572651b1eb 100644
--- a/packages/react-native-executorch/src/hooks/useOpticalCharacterRecognizer.ts
+++ b/packages/react-native-executorch/src/hooks/useOpticalCharacterRecognizer.ts
@@ -3,19 +3,24 @@ import { useResourceDownload, type ResourceOptions } from './useResourceDownload
 import { useModel } from './useModel';
 
 /**
- * React hook for the PP-OCRv6 pipeline. It downloads and loads the model, tracks
- * progress and errors, instantiates the task runner, and cleans up native memory
- * on unmount or config change. Heavy work runs on a worklet thread;
- * `recognizeCharacters` resolves with the recognized regions in reading order.
+ * React hook for the PP-OCRv6 pipeline.
+ *
+ * It downloads and loads the model, tracks progress and errors, instantiates the
+ * task runner, and cleans up native memory on unmount or config change. Heavy
+ * work runs on a worklet thread; `recognizeCharacters` resolves with the
+ * recognized regions in reading order.
+ *
+ * For imperative usage, see {@link createPaddleOcr}.
  * @category Hooks
  * @param config OCR model configuration. Use a preset from `models.ocr.*`.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns Readiness flags, download progress, and the `recognizeCharacters` /
- * `recognizeCharactersWorklet` runners.
+ * @returns The same object as {@link createPaddleOcr} (without `dispose`),
+ * combined with loading state, download progress, and resource info.
+ * @see {@link createPaddleOcr}
  */
 export function useOpticalCharacterRecognizer(config: PaddleOcrModel, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createPaddleOcr, resource ?? null);
+  const { model, error } = useModel(createPaddleOcr, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts b/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
index 88b182de07..6a0ed8872e 100644
--- a/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
+++ b/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
@@ -8,25 +8,28 @@ import {
 /**
  * React hook to load and run a privacy filter (PII detection) model.
  *
- * This hook manages downloading (if they are remote URLs) and loading both the
- * model file and its `tokenizer.json`, tracking download progress and errors,
- * and cleaning up native memory when the component unmounts or the
- * configuration changes.
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets and tokenizer files, tracking download progress and load errors,
+ * and releasing native memory when the component unmounts or the configuration
+ * changes.
+ *
+ * For imperative usage, see {@link createPrivacyFilter}.
  * @category Hooks
  * @typeParam Label The model's BIOES label space, narrowing the detected entity
  * types when a concrete `models` registry entry is passed.
- * @param config The privacy filter model configuration (model and tokenizer
- * paths plus the label space options).
+ * @param config The privacy filter model configuration. See {@link
+ * PrivacyFilterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and detection functions.
+ * @see {@link createPrivacyFilter}
  */
 export function usePrivacyFilter<Label extends string>(
   config: PrivacyFilterModel<Label>,
   options?: ResourceOptions
 ) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createPrivacyFilter<Label>, resource ?? null);
+  const { model, error } = useModel(createPrivacyFilter<Label>, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useResourceDownload.ts b/packages/react-native-executorch/src/hooks/useResourceDownload.ts
index 58df8b69aa..397fbe042b 100644
--- a/packages/react-native-executorch/src/hooks/useResourceDownload.ts
+++ b/packages/react-native-executorch/src/hooks/useResourceDownload.ts
@@ -7,7 +7,7 @@ import { isRnExecuTorchError } from '../core/error';
  * built on top of it.
  * @category Types
  */
-export interface ResourceOptions {
+export type ResourceOptions = {
   /** If true, prevents checks and downloads, resetting the hook state. */
   preventLoad?: boolean;
   /**
@@ -16,7 +16,7 @@ export interface ResourceOptions {
    * that changed behind a stable URL.
    */
   forceDownload?: boolean;
-}
+};
 
 /**
  * React hook to manage downloading and local caching of the remote resources
@@ -35,19 +35,23 @@ export interface ResourceOptions {
  * cheap for the files themselves (already-cached URLs are a no-op) but does
  * rebuild whatever pipeline consumes `resource` — correct, since `create<Task>`
  * factories bake their options in at construction time.
+ *
+ * For imperative usage, see {@link download}.
  * @category Hooks
  * @typeParam T The shape of the value being resolved.
- * @param config The value whose remote URLs should be resolved to local paths.
+ * @param config The value whose remote URLs should be resolved to local paths,
+ * or `undefined` to prevent loading.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the resolved value, the download progress
  * percentage, and any download error.
+ * @see {@link download}
  */
-export function useResourceDownload<T>(config: T, options?: ResourceOptions) {
+export function useResourceDownload<T>(config: T | undefined, options?: ResourceOptions) {
   const { preventLoad, forceDownload } = options ?? {};
 
   const [resource, setResource] = useState<T>();
   const [downloadProgress, setDownloadProgress] = useState(0);
-  const [downloadError, setDownloadError] = useState<Error | null>(null);
+  const [downloadError, setDownloadError] = useState<Error | undefined>();
 
   // Configs are plain JSON data, so serializing is a sound structural identity
   // and keeps an inline `config` object from re-running the effect every render.
@@ -56,9 +60,9 @@ export function useResourceDownload<T>(config: T, options?: ResourceOptions) {
   useEffect(() => {
     setResource(undefined);
     setDownloadProgress(0);
-    setDownloadError(null);
+    setDownloadError(undefined);
 
-    if (preventLoad) return;
+    if (config === undefined || preventLoad) return;
 
     let isMounted = true;
     const controller = new AbortController();
diff --git a/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts b/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
index 72220c5858..198c305758 100644
--- a/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
@@ -8,23 +8,27 @@ import {
 /**
  * React hook to load and run a semantic segmentation model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createSemanticSegmenter}.
  * @category Hooks
  * @typeParam L The type representing the segmentation labels.
- * @param config The semantic segmentation model configuration.
+ * @param config The semantic segmentation model configuration. See {@link
+ * SemanticSegmenterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
- * progress, and segmentation functions.
+ * progress, labels, and semantic segmentation functions.
+ * @see {@link createSemanticSegmenter}
  */
 export function useSemanticSegmenter<L extends PropertyKey = string>(
   config: SemanticSegmenterModel<L>,
   options?: ResourceOptions
 ) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createSemanticSegmenter<L>, resource ?? null);
+  const { model, error } = useModel(createSemanticSegmenter<L>, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useSpeechToText.ts b/packages/react-native-executorch/src/hooks/useSpeechToText.ts
index e2351ac47f..c493b34729 100644
--- a/packages/react-native-executorch/src/hooks/useSpeechToText.ts
+++ b/packages/react-native-executorch/src/hooks/useSpeechToText.ts
@@ -7,17 +7,21 @@ import {
 } from '../extensions/speech/tasks/whisperSpeechToText';
 
 /**
- * React hook to load and run a Whisper speech-to-text model.
+ * React hook to load and run a Whisper speech-to-text (ASR) model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model,
- * tokenizer, and FSMN-VAD, compiling them, tracking download progress and
- * compilation errors, and cleaning up native model memory when the component
- * unmounts or configuration changes.
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, tokenizer, and voice activity detector, tracking download
+ * progress and load errors, and releasing native memory when the component
+ * unmounts or the configuration changes.
+ *
+ * For imperative usage, see {@link createWhisperSpeechToText}.
  * @category Hooks
- * @param config The Whisper speech-to-text model configuration.
+ * @param config The Whisper speech-to-text model configuration. See {@link
+ * WhisperSttModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and transcription functions.
+ * @see {@link createWhisperSpeechToText}
  */
 export function useSpeechToText<L extends WhisperLanguage = WhisperLanguage>(
   config: WhisperSttModel<L>,
@@ -26,7 +30,7 @@ export function useSpeechToText<L extends WhisperLanguage = WhisperLanguage>(
   // Resolves the model, the tokenizer and the nested VAD model in one pass,
   // with progress weighted across all three.
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createWhisperSpeechToText, resource ?? null);
+  const { model, error } = useModel(createWhisperSpeechToText, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useStyleTransfer.ts b/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
index e853188b9a..ced118d0ad 100644
--- a/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
+++ b/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
@@ -5,19 +5,23 @@ import { createStyleTransfer, type StyleTransferModel } from '../extensions/cv/t
 /**
  * React hook to load and run an image style transfer model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createStyleTransfer}.
  * @category Hooks
- * @param config The style transfer model configuration.
+ * @param config The style transfer model configuration. See {@link
+ * StyleTransferModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and style transfer functions.
+ * @see {@link createStyleTransfer}
  */
 export function useStyleTransfer(config: StyleTransferModel, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createStyleTransfer, resource ?? null);
+  const { model, error } = useModel(createStyleTransfer, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useTextEmbedder.ts b/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
index ec93669a3b..498fbfda18 100644
--- a/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
+++ b/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
@@ -3,22 +3,25 @@ import { useResourceDownload, type ResourceOptions } from './useResourceDownload
 import { createTextEmbedder, type TextEmbedderModel } from '../extensions/nlp/tasks/textEmbedding';
 
 /**
- * React hook to load and run a text embedder model.
+ * React hook to load and run a text embedding model.
  *
- * This hook manages downloading (if they are remote URLs) and loading both the
- * model file and its `tokenizer.json`, tracking download progress and errors,
- * and cleaning up native memory when the component unmounts or the configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets and tokenizer files, tracking download progress and load errors,
+ * and releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createTextEmbedder}.
  * @category Hooks
- * @param config The text embedder model configuration (model and tokenizer
- * paths).
+ * @param config The text embedder model configuration. See {@link
+ * TextEmbedderModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and embedding functions.
+ * @see {@link createTextEmbedder}
  */
 export function useTextEmbedder(config: TextEmbedderModel, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createTextEmbedder, resource ?? null);
+  const { model, error } = useModel(createTextEmbedder, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useTextToImage.ts b/packages/react-native-executorch/src/hooks/useTextToImage.ts
index ff2676aa87..a4a5646ad5 100644
--- a/packages/react-native-executorch/src/hooks/useTextToImage.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToImage.ts
@@ -6,21 +6,24 @@ import {
 } from '../extensions/cv/tasks/sdxsTextToImage';
 
 /**
- * React hook to load and run the SDXS text-to-image pipeline.
+ * React hook to load and run an SDXS text-to-image synthesis model.
  *
- * It manages downloading (if the sources are remote URLs) and loading the
- * combined `.pte` program and the CLIP tokenizer, tracks download progress and
- * errors, and cleans up native memory when the component unmounts or the
- * configuration changes.
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets and tokenizer files, tracking download progress and load errors,
+ * and releasing native memory when the component unmounts or the configuration
+ * changes.
+ *
+ * For imperative usage, see {@link createSdxsTextToImage}.
  * @category Hooks
- * @param config The SDXS model configuration.
+ * @param config The SDXS model configuration. See {@link SdxsTextToImageModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
- * progress, and generation functions.
+ * progress, and image generation functions.
+ * @see {@link createSdxsTextToImage}
  */
 export function useTextToImage(config: SdxsTextToImageModel, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createSdxsTextToImage, resource ?? null);
+  const { model, error } = useModel(createSdxsTextToImage, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
index 2d6bfecc64..ddfce51d51 100644
--- a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
@@ -28,7 +28,7 @@ export type TtsHookResult<
   /** Whether the pipeline is loaded and ready to synthesize. */
   isReady: boolean;
   /** The download or load error, if any. */
-  error: Error | null;
+  error: Error | undefined;
   /** Download progress across every asset, in percent. */
   downloadProgress: number;
   /** The config with every remote URL resolved to a local path. */
@@ -40,17 +40,21 @@ export type TtsHookResult<
 };
 
 /**
- * React hook to load and manage the Kokoro Text-to-Speech pipeline.
+ * React hook to load and run the Kokoro Text-to-Speech pipeline.
  *
- * It manages downloading (if the sources are remote URLs) and loading the 2 sub-model
- * `.pte` files, the phonemizer assets and all voice `.bin` files, tracking download
- * progress and errors, and cleaning up native memory when unmounting.
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, phonemizer files, and voice style vectors, tracking download
+ * progress and load errors, and releasing native memory when the component
+ * unmounts or the configuration changes.
+ *
+ * For imperative usage, see {@link createKokoroTextToSpeech}.
  * @category Hooks
  * @typeParam K Voice keys record constraint.
- * @param config The Kokoro TTS model configuration.
+ * @param config The Kokoro TTS model configuration. See {@link KokoroTtsModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download progress,
- * and synthesis functions.
+ * @returns An object containing the model's loading state, error, download
+ * progress, and speech synthesis functions.
+ * @see {@link createKokoroTextToSpeech}
  */
 export function useTextToSpeech<K extends PropertyKey>(
   config: KokoroTtsModel<K>,
@@ -58,17 +62,22 @@ export function useTextToSpeech<K extends PropertyKey>(
 ): TtsHookResult<KokoroTtsModel<K>, KokoroTts<K>>;
 
 /**
- * React hook to load and manage the Supertonic 3 Text-to-Speech pipeline.
+ * React hook to load and run the Supertonic 3 Text-to-Speech pipeline.
+ *
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, indexer files, and voice style vectors, tracking download
+ * progress and load errors, and releasing native memory when the component
+ * unmounts or the configuration changes.
  *
- * It manages downloading (if the sources are remote URLs) and loading the 4 sub-model
- * `.pte` files, `unicode_indexer.json`, and all voice style `.json` files, tracking download
- * progress and errors, and cleaning up native memory when unmounting.
+ * For imperative usage, see {@link createSupertonicTextToSpeech}.
  * @category Hooks
  * @typeParam K Voice style keys record constraint.
- * @param config The Supertonic TTS model configuration.
+ * @param config The Supertonic TTS model configuration. See {@link
+ * SupertonicTtsModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download progress,
- * and synthesis functions.
+ * @returns An object containing the model's loading state, error, download
+ * progress, and speech synthesis functions.
+ * @see {@link createSupertonicTextToSpeech}
  */
 export function useTextToSpeech<K extends PropertyKey>(
   config: SupertonicTtsModel<K>,
@@ -88,7 +97,7 @@ export function useTextToSpeech<K extends PropertyKey>(
   ) => Promise<KokoroTts<K> | SupertonicTts<K>>;
 
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(create, resource ?? null);
+  const { model, error } = useModel(create, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useTokenizer.ts b/packages/react-native-executorch/src/hooks/useTokenizer.ts
index 4eceb71456..4c949e06d1 100644
--- a/packages/react-native-executorch/src/hooks/useTokenizer.ts
+++ b/packages/react-native-executorch/src/hooks/useTokenizer.ts
@@ -3,20 +3,24 @@ import { useResourceDownload, type ResourceOptions } from './useResourceDownload
 import { createTokenizer } from '../extensions/nlp/tasks/tokenization';
 
 /**
- * React hook to load and use a HuggingFace tokenizer.
+ * React hook to load and run a tokenizer.
  *
- * This hook manages downloading the `tokenizer.json` file (if it's a remote
- * URL), loading it natively, tracking download progress and load errors, and
- * cleaning up native memory when the component unmounts or the source changes.
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * tokenizer configuration files, tracking download progress and load errors,
+ * and releasing native memory when the component unmounts or the configuration
+ * changes.
+ *
+ * For imperative usage, see {@link createTokenizer}.
  * @category Hooks
  * @param tokenizerPath A remote URL or local path to a `tokenizer.json` file.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the tokenizer's loading state, error, download
  * progress, and tokenization functions.
+ * @see {@link createTokenizer}
  */
 export function useTokenizer(tokenizerPath: string, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(tokenizerPath, options);
-  const { model, error } = useModel(createTokenizer, resource ?? null);
+  const { model, error } = useModel(createTokenizer, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts b/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts
index bca07bb0dc..d0ef117b59 100644
--- a/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts
@@ -6,21 +6,24 @@ import {
 } from '../extensions/speech/tasks/fsmnVoiceActivityDetection';
 
 /**
- * React hook to load and run a Voice Activity Detection model.
+ * React hook to load and run a Voice Activity Detection (VAD) model.
  *
- * This hook manages downloading (if it's a remote URL) and loading the model
- * file, compiling it, tracking download progress and compilation errors, and
- * cleaning up native model memory when the component unmounts or configuration
+ * This hook manages downloading (if remote URLs are provided) and loading the
+ * model assets, compiling them, tracking download progress and load errors, and
+ * releasing native memory when the component unmounts or the configuration
  * changes.
+ *
+ * For imperative usage, see {@link createFsmnVoiceActivityDetector}.
  * @category Hooks
- * @param config The VAD model configuration.
+ * @param config The VAD model configuration. See {@link FsmnVadModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, one-shot detection functions, and live detection controls.
+ * @see {@link createFsmnVoiceActivityDetector}
  */
 export function useVoiceActivityDetector(config: FsmnVadModel, options?: ResourceOptions) {
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(createFsmnVoiceActivityDetector, resource ?? null);
+  const { model, error } = useModel(createFsmnVoiceActivityDetector, resource);
 
   return {
     isReady: !!model,
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index 278b223be1..4afa8bcb6e 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -55,9 +55,7 @@ export async function inspectModel(source: string): Promise<{
     model = loadModel(localPath);
     return { source, schema: model.schema, backends: model.backends };
   } finally {
-    if (model) {
-      model.dispose();
-    }
+    model?.dispose();
     if (downloaded) {
       await RNBlobUtil.fs.unlink(localPath).catch(() => {});
     }

From 249bb84a91285818b7094af996f22be079494429 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 15:29:43 +0200
Subject: [PATCH 03/38] docs(hooks): format JSDoc type links onto dedicated
 lines to prevent Prettier wrapping

---
 .../src/extensions/speech/tasks/whisperSpeechToText.ts        | 4 ++--
 packages/react-native-executorch/src/hooks/useClassifier.ts   | 4 ++--
 .../react-native-executorch/src/hooks/useImageEmbedder.ts     | 4 ++--
 .../react-native-executorch/src/hooks/useInstanceSegmenter.ts | 4 ++--
 .../react-native-executorch/src/hooks/useKeypointDetector.ts  | 4 ++--
 .../react-native-executorch/src/hooks/useLLMChatSession.ts    | 4 ++--
 .../react-native-executorch/src/hooks/useObjectDetector.ts    | 4 ++--
 .../react-native-executorch/src/hooks/usePrivacyFilter.ts     | 4 ++--
 .../react-native-executorch/src/hooks/useSemanticSegmenter.ts | 4 ++--
 packages/react-native-executorch/src/hooks/useSpeechToText.ts | 4 ++--
 .../react-native-executorch/src/hooks/useStyleTransfer.ts     | 4 ++--
 packages/react-native-executorch/src/hooks/useTextEmbedder.ts | 4 ++--
 packages/react-native-executorch/src/hooks/useTextToSpeech.ts | 4 ++--
 13 files changed, 26 insertions(+), 26 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
index 54fb6582df..01ed991102 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
@@ -98,8 +98,8 @@ export type WhisperSttModel<L extends WhisperLanguage = WhisperLanguage> = {
 /**
  * Loads a Whisper model and returns a set of transcription helpers.
  * @category Typescript API
- * @param config Model paths and supported-language metadata. See {@link
- * WhisperSttModel}.
+ * @param config Model paths and supported-language metadata.
+ * See {@link WhisperSttModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing transcription and
diff --git a/packages/react-native-executorch/src/hooks/useClassifier.ts b/packages/react-native-executorch/src/hooks/useClassifier.ts
index 795e010589..c129d7d0bc 100644
--- a/packages/react-native-executorch/src/hooks/useClassifier.ts
+++ b/packages/react-native-executorch/src/hooks/useClassifier.ts
@@ -13,8 +13,8 @@ import { createClassifier, type ClassifierModel } from '../extensions/cv/tasks/c
  * For imperative usage, see {@link createClassifier}.
  * @category Hooks
  * @typeParam L The type representing the classification labels.
- * @param config The image classification model configuration. See {@link
- * ClassifierModel}.
+ * @param config The image classification model configuration.
+ * See {@link ClassifierModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, labels, and classification functions.
diff --git a/packages/react-native-executorch/src/hooks/useImageEmbedder.ts b/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
index 5d91c0bfa2..f9669ee0a4 100644
--- a/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
+++ b/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
@@ -15,8 +15,8 @@ import {
  *
  * For imperative usage, see {@link createImageEmbedder}.
  * @category Hooks
- * @param config The image embedder model configuration. See {@link
- * ImageEmbedderModel}.
+ * @param config The image embedder model configuration.
+ * See {@link ImageEmbedderModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and embedding functions.
diff --git a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
index b43b6eba25..a9bd55bb7d 100644
--- a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
@@ -18,8 +18,8 @@ import {
  * @category Hooks
  * @typeParam F The bounding box format.
  * @typeParam L The class labels type.
- * @param config The instance segmentation model configuration. See {@link
- * InstanceSegmenterModel}.
+ * @param config The instance segmentation model configuration.
+ * See {@link InstanceSegmenterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, labels, and instance segmentation functions.
diff --git a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
index 01bba9c1c1..58ba581d97 100644
--- a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
@@ -18,8 +18,8 @@ import {
  * @category Hooks
  * @typeParam F The bounding box format.
  * @typeParam L The landmark labels type.
- * @param config The keypoint detection model configuration. See {@link
- * KeypointDetectorModel}.
+ * @param config The keypoint detection model configuration.
+ * See {@link KeypointDetectorModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, landmarks, and keypoint detection functions.
diff --git a/packages/react-native-executorch/src/hooks/useLLMChatSession.ts b/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
index c33a6a6c76..6b93393596 100644
--- a/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
+++ b/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
@@ -17,8 +17,8 @@ import {
  * For imperative usage, see {@link createLLMChatSession}.
  * @category Hooks
  * @param config The LLM model configuration. See {@link LLMModel}.
- * @param options Chat session options and load/caching options. See {@link
- * ResourceOptions}.
+ * @param options Chat session options and load/caching options.
+ * See {@link ResourceOptions}.
  * @returns An object containing the session's loading state, error, download
  * progress, and chat functions.
  * @see {@link createLLMChatSession}
diff --git a/packages/react-native-executorch/src/hooks/useObjectDetector.ts b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
index e66109628a..17241767b1 100644
--- a/packages/react-native-executorch/src/hooks/useObjectDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
@@ -18,8 +18,8 @@ import {
  * @category Hooks
  * @typeParam F The bounding box format.
  * @typeParam L The type representing the object class labels.
- * @param config The object detection model configuration. See {@link
- * ObjectDetectorModel}.
+ * @param config The object detection model configuration.
+ * See {@link ObjectDetectorModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, labels, and object detection functions.
diff --git a/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts b/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
index 6a0ed8872e..47a567408b 100644
--- a/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
+++ b/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
@@ -17,8 +17,8 @@ import {
  * @category Hooks
  * @typeParam Label The model's BIOES label space, narrowing the detected entity
  * types when a concrete `models` registry entry is passed.
- * @param config The privacy filter model configuration. See {@link
- * PrivacyFilterModel}.
+ * @param config The privacy filter model configuration.
+ * See {@link PrivacyFilterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and detection functions.
diff --git a/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts b/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
index 198c305758..c37800ca30 100644
--- a/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
@@ -16,8 +16,8 @@ import {
  * For imperative usage, see {@link createSemanticSegmenter}.
  * @category Hooks
  * @typeParam L The type representing the segmentation labels.
- * @param config The semantic segmentation model configuration. See {@link
- * SemanticSegmenterModel}.
+ * @param config The semantic segmentation model configuration.
+ * See {@link SemanticSegmenterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, labels, and semantic segmentation functions.
diff --git a/packages/react-native-executorch/src/hooks/useSpeechToText.ts b/packages/react-native-executorch/src/hooks/useSpeechToText.ts
index c493b34729..75380a9d3c 100644
--- a/packages/react-native-executorch/src/hooks/useSpeechToText.ts
+++ b/packages/react-native-executorch/src/hooks/useSpeechToText.ts
@@ -16,8 +16,8 @@ import {
  *
  * For imperative usage, see {@link createWhisperSpeechToText}.
  * @category Hooks
- * @param config The Whisper speech-to-text model configuration. See {@link
- * WhisperSttModel}.
+ * @param config The Whisper speech-to-text model configuration.
+ * See {@link WhisperSttModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and transcription functions.
diff --git a/packages/react-native-executorch/src/hooks/useStyleTransfer.ts b/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
index ced118d0ad..e9780cb6bd 100644
--- a/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
+++ b/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
@@ -12,8 +12,8 @@ import { createStyleTransfer, type StyleTransferModel } from '../extensions/cv/t
  *
  * For imperative usage, see {@link createStyleTransfer}.
  * @category Hooks
- * @param config The style transfer model configuration. See {@link
- * StyleTransferModel}.
+ * @param config The style transfer model configuration.
+ * See {@link StyleTransferModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and style transfer functions.
diff --git a/packages/react-native-executorch/src/hooks/useTextEmbedder.ts b/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
index 498fbfda18..779980ec9e 100644
--- a/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
+++ b/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
@@ -12,8 +12,8 @@ import { createTextEmbedder, type TextEmbedderModel } from '../extensions/nlp/ta
  *
  * For imperative usage, see {@link createTextEmbedder}.
  * @category Hooks
- * @param config The text embedder model configuration. See {@link
- * TextEmbedderModel}.
+ * @param config The text embedder model configuration.
+ * See {@link TextEmbedderModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and embedding functions.
diff --git a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
index ddfce51d51..47d5814a5a 100644
--- a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
@@ -72,8 +72,8 @@ export function useTextToSpeech<K extends PropertyKey>(
  * For imperative usage, see {@link createSupertonicTextToSpeech}.
  * @category Hooks
  * @typeParam K Voice style keys record constraint.
- * @param config The Supertonic TTS model configuration. See {@link
- * SupertonicTtsModel}.
+ * @param config The Supertonic TTS model configuration.
+ * See {@link SupertonicTtsModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
  * @returns An object containing the model's loading state, error, download
  * progress, and speech synthesis functions.

From 11f25eaab284955152b2cf26d0ba178c187e45be Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 15:53:03 +0200
Subject: [PATCH 04/38] docs: add top-level module documentation to index,
 models, constants, and utils

---
 .../react-native-executorch/src/constants.ts  |  8 ++++
 packages/react-native-executorch/src/index.ts | 46 +++++++++++++++++++
 .../react-native-executorch/src/models.ts     | 10 ++++
 packages/react-native-executorch/src/utils.ts |  6 +++
 4 files changed, 70 insertions(+)

diff --git a/packages/react-native-executorch/src/constants.ts b/packages/react-native-executorch/src/constants.ts
index 0bb489bfa0..d5dd31327c 100644
--- a/packages/react-native-executorch/src/constants.ts
+++ b/packages/react-native-executorch/src/constants.ts
@@ -1,3 +1,11 @@
+/**
+ * Shared constants, label maps, and default configurations for task pipelines.
+ *
+ * This module exports constant arrays, dataset labels, and normalization
+ * parameters used across task pipelines and model configurations.
+ * @module Constants
+ */
+
 /**
  * ImageNet 1K dataset label array containing the 1000 categories.
  * @category Constants
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 920ff8344c..9fc98cd7b1 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -1,3 +1,49 @@
+/**
+ * React Native ExecuTorch is a high-performance, privacy-first on-device AI
+ * inference library for React Native, powered by PyTorch's ExecuTorch runtime.
+ *
+ * It provides a layered architecture designed for different use cases:
+ *
+ * - **React Hooks (`use<Task>`)**:
+ *   Declarative hooks (e.g. {@link useClassifier}, {@link useLLMChatSession},
+ *   {@link useSpeechToText}) designed for UI components. They manage
+ *   downloading remote model assets with progress tracking, compilation, and
+ *   automatic native memory disposal on unmount.
+ *
+ * - **Task APIs (`create<Task>`)**:
+ *   Imperative, promise-based pipelines (e.g. {@link createClassifier},
+ *   {@link createLLMChatSession}, {@link createWhisperSpeechToText}) with
+ *   built-in preprocessing, postprocessing, and tokenizers. Ideal for
+ *   background services, worklets, or apps needing manual lifecycle control.
+ *
+ * - **Domain Utilities (`cv`, `llm`, `speech`, `nlp`, `math`)**:
+ *   Domain-specific low-level tools:
+ *   - `cv`: Image transformations (resizing, normalization, layout conversions),
+ *     bounding box and keypoint helpers.
+ *   - `llm`: Token-by-token text generation runner, chat template formatting,
+ *     and function calling helpers.
+ *   - `nlp`: Fast native HuggingFace tokenizers and privacy filter utilities.
+ *   - `speech`: Text-to-speech phonemizers, sentence splitters, voice activity
+ *     detection, and audio utilities.
+ *   - `math`: Native C++ tensor operations (softmax, sigmoid, argmax, linear,
+ *     activations).
+ *
+ * - **Model Registry ({@link models})**:
+ *   A curated catalog of verified, hosted on-device AI models across LLMs,
+ *   Computer Vision, Speech, and NLP. Provides download URLs, pre-tuned
+ *   configurations, and label maps for out-of-the-box inference.
+ *
+ * - **Resource Fetcher ({@link download})**:
+ *   Imperative asset downloader and caching engine with abort control, progress
+ *   tracking, and deduplicated local storage.
+ *
+ * - **Core Primitives ({@link Model}, {@link Tensor}, {@link schema}, {@link wrapAsync})**:
+ *   Low-level building blocks for custom architectures: direct C++ tensor
+ *   memory management, raw model execution, load-time shape/domain validation,
+ *   and worklet threading.
+ * @packageDocumentation
+ */
+
 // Hooks — primary API for app developers
 export * from './hooks/useClassifier';
 export * from './hooks/useStyleTransfer';
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 72ea6530f2..690aae6322 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -1,3 +1,13 @@
+/**
+ * Pre-configured registry of hosted on-device AI models.
+ *
+ * Provides ready-to-use configurations for popular models across computer
+ * vision, speech synthesis/recognition, natural language processing, and large
+ * language models (LLMs). Each entry includes verified remote `.pte` download
+ * URLs, tokenizer/phonemizer files, preprocessing parameters, and label maps.
+ * @module Models
+ */
+
 import type { ClassifierModel } from './extensions/cv/tasks/classification';
 import type { ObjectDetectorModel } from './extensions/cv/tasks/objectDetection';
 import type { StyleTransferModel } from './extensions/cv/tasks/styleTransfer';
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index 4afa8bcb6e..24e3cc7652 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -1,3 +1,9 @@
+/**
+ * Utility functions for inspecting models and querying native runtime
+ * capabilities.
+ * @module Utils
+ */
+
 import { rnexecutorchJsi } from './native/bridge';
 import { loadModel } from './core/model';
 import type { ModelSpec, ConcreteDim } from './core/schema';

From c370798362f20578003cf19f8da98a331f7d14bc Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 16:00:50 +0200
Subject: [PATCH 05/38] refactor(hooks): use switch statement for TTS factory
 resolution in useTextToSpeech

---
 .../src/hooks/useTextToSpeech.ts                  | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
index 47d5814a5a..f94f6ea030 100644
--- a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
@@ -90,14 +90,19 @@ export function useTextToSpeech<K extends PropertyKey>(
 ) {
   // Each config names the pipeline it belongs to, so the factory is resolved
   // from that tag alone.
-  const create = (
-    config.name === 'kokoro' ? createKokoroTextToSpeech : createSupertonicTextToSpeech
-  ) as (
+  const factory = (
     ttsConfig: KokoroTtsModel<K> | SupertonicTtsModel<K>
-  ) => Promise<KokoroTts<K> | SupertonicTts<K>>;
+  ): Promise<KokoroTts<K> | SupertonicTts<K>> => {
+    switch (ttsConfig.name) {
+      case 'kokoro':
+        return createKokoroTextToSpeech(ttsConfig);
+      case 'supertonic':
+        return createSupertonicTextToSpeech(ttsConfig);
+    }
+  };
 
   const { resource, downloadProgress, downloadError } = useResourceDownload(config, options);
-  const { model, error } = useModel(create, resource);
+  const { model, error } = useModel(factory, resource);
 
   return {
     isReady: !!model,

From dd2700cca6ef352153d94d8c404be60c61d44c35 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 16:11:36 +0200
Subject: [PATCH 06/38] docs(core): improve error documentation and linking

---
 .../react-native-executorch/src/core/error.ts | 46 ++++++++++++-------
 1 file changed, 30 insertions(+), 16 deletions(-)

diff --git a/packages/react-native-executorch/src/core/error.ts b/packages/react-native-executorch/src/core/error.ts
index 048e722e91..74601124e0 100644
--- a/packages/react-native-executorch/src/core/error.ts
+++ b/packages/react-native-executorch/src/core/error.ts
@@ -1,14 +1,14 @@
 /**
- * Errors raised by React Native ExecuTorch.
+ * Classified error handling for React Native ExecuTorch.
  *
- * This module is the source of truth for the error contract; `cpp/core/error.h`
- * mirrors it by hand, the same way the rest of the TS/JSI interface is mirrored.
+ * All errors raised by the library carry a machine-readable
+ * {@link RnExecuTorchErrorCode} (e.g. `LOAD_FAILED`, `EXECUTION_FAILED`,
+ * `RESOURCE_BUSY`, or `SCHEMA_MISMATCH`) allowing applications to inspect and
+ * handle failures programmatically.
  *
- * Errors are plain `Error` objects with extra fields rather than a class.
- * Worklet runtimes are separate JavaScript runtimes and a value thrown on one
- * does not keep its class identity or prototype chain when it travels to
- * another, so a class would only work on some of the paths that can throw.
- * @packageDocumentation
+ * Use {@link isRnExecuTorchError} to safely narrow caught errors across
+ * asynchronous calls and worklet runtime boundaries.
+ * @module Core/Error
  */
 
 /**
@@ -43,8 +43,17 @@ export const VALID_ERROR_CODES = [
 export type RnExecuTorchErrorCode = (typeof VALID_ERROR_CODES)[number];
 
 /**
- * An error raised by React Native ExecuTorch: a standard `Error` carrying a
- * {@link RnExecuTorchErrorCode}.
+ * An error raised by React Native ExecuTorch.
+ *
+ * Represents a standard `Error` augmented with a machine-readable
+ * {@link RnExecuTorchErrorCode} and an optional ExecuTorch C++ runtime code
+ * (`etRuntimeErrorCode`).
+ *
+ * When thrown, call as a factory function without `new` (safe across worklet
+ * threads):
+ * ```typescript
+ * throw RnExecuTorchError('INVALID_ARGUMENT', 'Shape dimensions must be positive');
+ * ```
  * @category Errors
  * @typeParam C The specific code, narrowed by {@link isRnExecuTorchError}.
  */
@@ -60,13 +69,18 @@ export type RnExecuTorchError<C extends RnExecuTorchErrorCode = RnExecuTorchErro
 };
 
 /**
- * Builds an {@link RnExecuTorchError}. Safe to call from anywhere, including
- * inside a worklet.
+ * Creates an {@link RnExecuTorchError} carrying a machine-readable code.
  * @category Errors
- * @param code The classification to attach.
- * @param message A human-readable description. Include the offending values.
- * @param etRuntimeErrorCode The raw ExecuTorch runtime error, when there is one.
- * @returns An `Error` carrying `code`.
+ * @typeParam C The specific error code.
+ * @param code The classification to attach. See {@link RnExecuTorchErrorCode}.
+ * @param message A human-readable description of the failure.
+ * @param etRuntimeErrorCode The raw ExecuTorch runtime error, when available.
+ * @returns An `Error` carrying `code` and `name: 'RnExecuTorchError'`.
+ * @see {@link RnExecuTorchErrorCode}
+ * @example
+ * ```typescript
+ * throw RnExecuTorchError('INVALID_ARGUMENT', 'Shape dimensions must be positive');
+ * ```
  */
 export function RnExecuTorchError<C extends RnExecuTorchErrorCode>(
   code: C,

From e6824aa2e1d59508a15cf955247892e2567fc550 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 16:16:04 +0200
Subject: [PATCH 07/38] docs(core): improve model and tensor documentation and
 error codes

---
 .../react-native-executorch/src/core/model.ts | 38 +++++++++++++++++--
 .../src/core/tensor.ts                        | 32 ++++++++++++++++
 2 files changed, 66 insertions(+), 4 deletions(-)

diff --git a/packages/react-native-executorch/src/core/model.ts b/packages/react-native-executorch/src/core/model.ts
index 5fd59bb2d4..867369fc1c 100644
--- a/packages/react-native-executorch/src/core/model.ts
+++ b/packages/react-native-executorch/src/core/model.ts
@@ -1,3 +1,12 @@
+/**
+ * Low-level ExecuTorch model loading, execution, and lifetime management.
+ *
+ * Provides direct access to compiled `.pte` models in native C++ memory. Loaded
+ * models expose synchronous method execution (`execute`) with pre-allocated
+ * output buffers and manual native memory cleanup (`dispose`).
+ * @module Core/Model
+ */
+
 import { rnexecutorchJsi } from '../native/bridge';
 import type { Tensor } from './tensor';
 import type { ModelSpec, ConcreteDim } from './schema';
@@ -24,10 +33,10 @@ export type ModelOutput = Tensor | number | boolean | string | null;
  * this interface.
  *
  * Obtain a `Model` instance via the {@link loadModel} function. When the model
- * is no longer needed call {@link Model.dispose} to release native memory.
+ * is no longer needed, call {@link Model.dispose} to release native memory.
  * @category Types
  */
-export interface Model {
+export type Model = {
   /** The local filesystem path of the `.pte` model file. */
   readonly path: string;
   /** The exported schema of this model. */
@@ -46,7 +55,11 @@ export interface Model {
    * @param inputs The list of input values to pass to the method, in order.
    * @param outputTensors Pre-allocated tensors for the method to write outputs
    * into, in order.
-   * @returns The list of output values produced by the method, in order.
+   * @throws {RnExecuTorchError} Thrown with code `EXECUTION_FAILED` if
+   * inference fails, `SCHEMA_MISMATCH` if runtime constraints fail,
+   * `RESOURCE_BUSY` if the model or a tensor is in use, `RESOURCE_DISPOSED` if
+   * disposed, or `INVALID_ARGUMENT` if inputs or output placeholders are
+   * invalid.
    */
   execute(methodName: string, inputs: ModelInput[], outputTensors: Tensor[]): ModelOutput[];
 
@@ -62,7 +75,7 @@ export interface Model {
    * @internal
    */
   readonly [modelBrand]: never;
-}
+};
 
 /**
  * Loads and compiles an ExecuTorch `.pte` model from the local filesystem.
@@ -73,6 +86,23 @@ export interface Model {
  * @category Typescript API
  * @param modelPath The absolute local path to the `.pte` model file.
  * @returns The compiled {@link Model} instance, ready for execution.
+ * @throws {RnExecuTorchError} Thrown with code `LOAD_FAILED` if the model file
+ * cannot be opened, has an invalid format, or fails native initialization.
+ * @see {@link wrapAsync}
+ * @example
+ * ```typescript
+ * const model = loadModel('/path/to/model.pte');
+ * const input = tensor('float32', [1, 3, 224, 224]);
+ * const output = tensor('float32', [1, 1000]);
+ * try {
+ *   model.execute('forward', [input], [output]);
+ *   // ...
+ * } finally {
+ *   input.dispose();
+ *   output.dispose();
+ *   model.dispose();
+ * }
+ * ```
  */
 export function loadModel(modelPath: string): Model {
   'worklet';
diff --git a/packages/react-native-executorch/src/core/tensor.ts b/packages/react-native-executorch/src/core/tensor.ts
index b54df997c5..5d7649accf 100644
--- a/packages/react-native-executorch/src/core/tensor.ts
+++ b/packages/react-native-executorch/src/core/tensor.ts
@@ -1,3 +1,12 @@
+/**
+ * Native C++ tensor allocation, data transfers, and memory management.
+ *
+ * Tensors are the fundamental data structures used throughout React Native
+ * ExecuTorch. They hold multidimensional typed arrays allocated in native
+ * heap memory and provide transfers to and from JavaScript typed arrays.
+ * @module Core/Tensor
+ */
+
 import { rnexecutorchJsi } from '../native/bridge';
 
 declare const tensorBrand: unique symbol;
@@ -37,6 +46,10 @@ export type Tensor = {
    * `numel - offset`, i.e. copies from `offset` to the end of the source
    * tensor.
    * @returns The destination tensor `dst`.
+   * @throws {RnExecuTorchError} Thrown with code `INVALID_ARGUMENT` if the copy
+   * bounds exceed the tensor size or data types mismatch, `RESOURCE_BUSY` if
+   * either tensor is in use, or `RESOURCE_DISPOSED` if either tensor was
+   * disposed.
    */
   copyTo(dst: Tensor, options?: { offset?: number; length?: number }): Tensor;
 
@@ -53,6 +66,9 @@ export type Tensor = {
    * tensor's size. Use a `BigInt64Array` for `int64` tensors and a
    * `Uint8Array` for `bool` tensors.
    * @returns `this` tensor.
+   * @throws {RnExecuTorchError} Thrown with code `INVALID_ARGUMENT` if `src`
+   * byte length does not match tensor size, `RESOURCE_BUSY` if the tensor is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   setData(src: Float32Array | Uint8Array | Int32Array | BigInt64Array): Tensor;
 
@@ -62,6 +78,9 @@ export type Tensor = {
    * @param dst The destination typed array. Its size in bytes must match
    * tensor's size.
    * @returns The same `dst` array, now filled with tensor data.
+   * @throws {RnExecuTorchError} Thrown with code `INVALID_ARGUMENT` if `dst`
+   * byte length does not match tensor size, `RESOURCE_BUSY` if the tensor is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   getData<T extends Float32Array | Uint8Array | Int32Array | BigInt64Array>(dst: T): T;
 
@@ -112,6 +131,19 @@ export type Tensor = {
  * @param src Optional typed array used to initialize the tensor's data. Its
  * size in bytes must match tensor's size.
  * @returns A newly allocated native tensor.
+ * @throws {RnExecuTorchError} Thrown with code `INVALID_ARGUMENT` if any
+ * dimension in `shape` is non-positive or if `src` byte length does not match
+ * the allocated tensor size.
+ * @example
+ * ```typescript
+ * const t = tensor('float32', [1, 4], new Float32Array([1.0, 2.0, 3.0, 4.0]));
+ * try {
+ *   const data = t.getData(new Float32Array(4));
+ *   console.log(data); // Float32Array [1, 2, 3, 4]
+ * } finally {
+ *   t.dispose();
+ * }
+ * ```
  */
 export function tensor(
   dtype: DType,

From e43597eeeed890a90ace11755a2d17038d13fc6e Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 16:16:57 +0200
Subject: [PATCH 08/38] docs(core): improve runtime documentation and examples

---
 .../react-native-executorch/src/core/runtime.ts  | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/packages/react-native-executorch/src/core/runtime.ts b/packages/react-native-executorch/src/core/runtime.ts
index d4f4335003..6c1f349061 100644
--- a/packages/react-native-executorch/src/core/runtime.ts
+++ b/packages/react-native-executorch/src/core/runtime.ts
@@ -1,3 +1,12 @@
+/**
+ * Background worklet execution and thread runtime management.
+ *
+ * Provides utilities to dispatch synchronous, heavy native operations (model
+ * compilation, tensor inference) onto dedicated background worklet threads,
+ * preventing them from blocking the React Native JavaScript thread.
+ * @module Core/Runtime
+ */
+
 import {
   createWorkletRuntime,
   runOnRuntimeAsync,
@@ -35,6 +44,13 @@ export const defaultWorkletRuntime = createWorkletRuntime({
  * {@link defaultWorkletRuntime}.
  * @returns An async function with the same signature as `fn` that resolves to
  * `fn`'s return value or rejects with an `RnExecuTorchError` if `fn` throws.
+ * @throws {RnExecuTorchError} Propagates any error thrown inside `fn` across the
+ * worklet thread boundary.
+ * @example
+ * ```typescript
+ * const asyncLoadModel = wrapAsync(loadModel);
+ * const model = await asyncLoadModel('/path/to/model.pte');
+ * ```
  */
 export function wrapAsync<Args extends any[], R>(
   fn: (...args: Args) => R,

From 80337d2c45a4fd876f233a6b8d5541cfce58f27b Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 16:17:39 +0200
Subject: [PATCH 09/38] docs(core): add module doc to schema

---
 .../src/core/schema.ts                        | 80 +++++++++----------
 1 file changed, 38 insertions(+), 42 deletions(-)

diff --git a/packages/react-native-executorch/src/core/schema.ts b/packages/react-native-executorch/src/core/schema.ts
index 446e5f615c..c0c2d9b13b 100644
--- a/packages/react-native-executorch/src/core/schema.ts
+++ b/packages/react-native-executorch/src/core/schema.ts
@@ -2,75 +2,71 @@
  * Model specs and spec validation.
  *
  * A model spec is a structural contract describing a model's methods: the
- * parameter specs of every input and output (primitive tags, tensor data
- * types, and per-dimension domains) and the runtime constraints the method
- * declares over its tensor dimensions. A spec is either:
+ * parameter specs of every input and output (primitive tags, tensor data types,
+ * and per-dimension domains) and the runtime constraints the method declares
+ * over its tensor dimensions. A spec is either:
  * - **allowed** (`SymbolicDim`) — written by a pipeline to state which models
- *   it can work with. Dimensions may be named symbols: `static` symbols bind
- *   to constant dimensions, `dynamic` symbols to ranges or enums, and reusing
- *   a symbol requires every occurrence to bind to the same domain. Several
+ *   it can work with. Dimensions may be named symbols: `static` symbols bind to
+ *   constant dimensions, `dynamic` symbols to ranges or enums, and reusing a
+ *   symbol requires every occurrence to bind to the same domain. Several
  *   allowed specs can be passed as variants; matching any one of them is
  *   enough.
  * - **exported** (`ConcreteDim`) — derived from an exported model's metadata,
  *   stating what the model actually provides.
  *
  * **Dimension domains vs runtime values.** The central distinction of this
- * module is between a dimension's *domain* and its *runtime value*. The
- * domain is the set of values the dimension may take: `constant` is a
- * singleton (the value is fully known statically), while `range` and `enum`
- * are proper sets that only narrow the possibilities. The runtime value is
- * the actual size of the dimension for a concrete tensor in one given
- * execution — a single element drawn from the domain.
+ * module is between a dimension's *domain* and its *runtime value*. The domain
+ * is the set of values the dimension may take: `constant` is a singleton (the
+ * value is fully known statically), while `range` and `enum` are proper sets
+ * that only narrow the possibilities. The runtime value is the actual size of
+ * the dimension for a concrete tensor in one given execution — a single element
+ * drawn from the domain.
  *
  * Accordingly, two different validations must not be confused:
  * - **Spec validation** (this module, {@link validateSpec}) — a static,
  *   load-time check that an exported spec satisfies an allowed spec. It only
  *   ever compares domains: symbols bind to domains (repeated symbols to the
  *   same one), and constraints are matched as declarations. Domain equality
- *   says nothing about runtime values — two dimensions with the same domain
- *   may still take different values in an execution.
- * - **Runtime validation** (the native runtime, not this module) — at
- *   execution time the runtime checks that every concrete tensor shape lies
- *   within the declared domains and enforces the declared runtime
- *   constraints.
+ *   says nothing about runtime values — two dimensions with the same domain may
+ *   still take different values in an execution.
+ * - **Runtime validation** (the native runtime, not this module) — at execution
+ *   time the runtime checks that every concrete tensor shape lies within the
+ *   declared domains and enforces the declared runtime constraints.
  *
  * Runtime constraints are statements about runtime values, not domains: an
- * equality constraint requires its dimensions' values to coincide in any
- * given execution, and a linear constraint requires them to satisfy
- * `lhs = a * rhs + b`. Since spec validation only sees domains, it cannot
- * decide whether such a relation holds — the exported spec must simply
- * declare exactly the allowed spec's constraints (1-to-1, no missing, no
- * extras). The only exception is degenerate: a constant domain has exactly
- * one possible value, so an equality constraint between constants is fully
- * decided statically — equal constants are equal at runtime. Linear
- * constraints, in contrast, are never evaluated against domains here, even
- * between constants.
+ * equality constraint requires its dimensions' values to coincide in any given
+ * execution, and a linear constraint requires them to satisfy `lhs = a * rhs +
+ * b`. Since spec validation only sees domains, it cannot decide whether such a
+ * relation holds — the exported spec must simply declare exactly the allowed
+ * spec's constraints (1-to-1, no missing, no extras). The only exception is
+ * degenerate: a constant domain has exactly one possible value, so an equality
+ * constraint between constants is fully decided statically — equal constants
+ * are equal at runtime. Linear constraints, in contrast, are never evaluated
+ * against domains here, even between constants.
  *
- * **Exported spec source.** An exported model's `ModelSpec<ConcreteDim>`
- * is populated at load time from one of two sources:
+ * **Exported spec source.** An exported model's `ModelSpec<ConcreteDim>` is
+ * populated at load time from one of two sources:
  *
  * 1. **ExecuTorch `MethodMeta`** (default) — when the `.pte` only carries
- *    static metadata, every dimension domain is `constant`. This is
- *    sufficient for models whose input/output shapes are fully fixed at
- *    export time.
+ *    static metadata, every dimension domain is `constant`. This is sufficient
+ *    for models whose input/output shapes are fully fixed at export time.
  *
- * 2. **Companion `get_model_schema` method** — for models whose tensors
- *    have dynamic or enumerated dimensions (e.g. variable-length sequences)
- *    or that declare runtime constraints, the `.pte` exports a method named
+ * 2. **Companion `get_model_schema` method** — for models whose tensors have
+ *    dynamic or enumerated dimensions (e.g. variable-length sequences) or that
+ *    declare runtime constraints, the `.pte` exports a method named
  *    `get_model_schema` that returns a JSON-encoded `ModelSpec<ConcreteDim>`
  *    string. The native loader calls this method after loading the model and
- *    merges the result into `model.schema`, overlaying precise `range`,
- *    `enum`, and `RuntimeConstraint` entries onto the base `MethodMeta`.
- *    Only methods that actually need overrides need to appear in the JSON;
- *    methods absent from the companion spec are kept as-is from
- *    `MethodMeta`.
+ *    merges the result into `model.schema`, overlaying precise `range`, `enum`,
+ *    and `RuntimeConstraint` entries onto the base `MethodMeta`. Only methods
+ *    that actually need overrides need to appear in the JSON; methods absent
+ *    from the companion spec are kept as-is from `MethodMeta`.
  *
  *    **Tip for model export:**
  *    Embed the companion schema method during ExecuTorch compilation in Python
  *    by passing `constant_methods={"get_model_schema": schema_json}` when
  *    lowering with `to_edge_transform_and_lower(...)` where `schema_json` is
  *    the JSON string encoding the model's `ModelSpec<ConcreteDim>`.
- * @packageDocumentation
+ * @module Core/Schema
  */
 import type { DType } from './tensor';
 import { RnExecuTorchError } from './error';

From 37c779167396bd4e046ee44febd11ffab322b8a6 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 16:23:28 +0200
Subject: [PATCH 10/38] fix(core): return raw ConcreteDim for untyped dim
 accessors and mark getRegisteredBackends as worklet

---
 .../src/core/schema.ts                        | 110 +++++++++++-------
 packages/react-native-executorch/src/utils.ts |   1 +
 2 files changed, 68 insertions(+), 43 deletions(-)

diff --git a/packages/react-native-executorch/src/core/schema.ts b/packages/react-native-executorch/src/core/schema.ts
index c0c2d9b13b..8cd52d9bc5 100644
--- a/packages/react-native-executorch/src/core/schema.ts
+++ b/packages/react-native-executorch/src/core/schema.ts
@@ -806,6 +806,72 @@ export type SpecMatch<K extends PropertyKey = PropertyKey> = {
   };
 };
 
+function resolveSymbolDim(bindings: SymbolBindings, name: string, kind?: string): any {
+  const dim = bindings.get(name);
+  if (!dim) {
+    throw RnExecuTorchError('INVALID_ARGUMENT', `Symbol '${name}' not found in bindings.`);
+  }
+  if (kind === 'constant') {
+    if (dim.kind !== 'constant') {
+      throw RnExecuTorchError(
+        'INVALID_ARGUMENT',
+        `Symbol '${name}' is '${dim.kind}', expected 'constant'.`
+      );
+    }
+    return dim.value;
+  }
+  if (kind === 'range') {
+    if (dim.kind !== 'range') {
+      throw RnExecuTorchError(
+        'INVALID_ARGUMENT',
+        `Symbol '${name}' is '${dim.kind}', expected 'range'.`
+      );
+    }
+    return dim.range;
+  }
+  if (kind === 'enum') {
+    if (dim.kind !== 'enum') {
+      throw RnExecuTorchError(
+        'INVALID_ARGUMENT',
+        `Symbol '${name}' is '${dim.kind}', expected 'enum'.`
+      );
+    }
+    return dim.choices;
+  }
+  if (kind === 'dynamic') {
+    if (dim.kind === 'constant') {
+      throw RnExecuTorchError(
+        'INVALID_ARGUMENT',
+        `Symbol '${name}' is 'constant', expected 'dynamic'.`
+      );
+    }
+    return dim;
+  }
+  return dim;
+}
+
+function createSpecMatch<K extends PropertyKey>(
+  variant: K,
+  bindings: SymbolBindings
+): SpecMatch<K> {
+  const dim: any = (name: string, kind?: string) => resolveSymbolDim(bindings, name, kind);
+  const createAccessor = (kind?: string) => {
+    return (...names: string[]): any => names.map((name) => resolveSymbolDim(bindings, name, kind));
+  };
+
+  return {
+    variant,
+    dim,
+    dims: {
+      any: createAccessor(),
+      enum: createAccessor('enum'),
+      range: createAccessor('range'),
+      dynamic: createAccessor('dynamic'),
+      constant: createAccessor('constant'),
+    },
+  };
+}
+
 /**
  * Validates that an exported (concrete) model spec satisfies at least one of
  * the allowed (symbolic) model specs — variants are tried in order and the
@@ -848,49 +914,7 @@ export function validateSpec<const T extends Record<string, ModelSpec<SymbolicDi
       matchModelSpecsSymbols(allowedModelSpec, exportedModelSpec, bindings);
       matchRuntimeConstraints(allowedModelSpec, exportedModelSpec);
 
-      const dimFn = (name: string, kind?: string): any => {
-        const dim = bindings.get(name);
-        if (!dim) {
-          throw RnExecuTorchError('INVALID_ARGUMENT', `Symbol '${name}' not found in bindings.`);
-        }
-        if (kind) {
-          if (kind === 'dynamic') {
-            if (dim.kind === 'constant') {
-              throw RnExecuTorchError(
-                'INVALID_ARGUMENT',
-                `Symbol '${name}' is 'constant', expected 'dynamic'.`
-              );
-            }
-            return dim;
-          }
-          if (dim.kind !== kind) {
-            throw RnExecuTorchError(
-              'INVALID_ARGUMENT',
-              `Symbol '${name}' is '${dim.kind}', expected '${kind}'.`
-            );
-          }
-        }
-        if (dim.kind === 'constant') return dim.value;
-        if (dim.kind === 'range') return dim.range;
-        if (dim.kind === 'enum') return dim.choices;
-        return dim;
-      };
-
-      const createAccessor = (kind?: string) => {
-        return (...names: string[]): any => names.map((name) => dimFn(name, kind));
-      };
-
-      return {
-        variant: key,
-        dim: dimFn,
-        dims: {
-          any: createAccessor(),
-          enum: createAccessor('enum'),
-          range: createAccessor('range'),
-          dynamic: createAccessor('dynamic'),
-          constant: createAccessor('constant'),
-        },
-      };
+      return createSpecMatch(key, bindings);
     } catch (e: any) {
       errors.push(`Variant '${key}': ${e.message}`);
       continue;
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index 24e3cc7652..12413b981a 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -17,6 +17,7 @@ import RNBlobUtil from 'react-native-blob-util';
  * 'CoreMLBackend').
  */
 export function getRegisteredBackends(): string[] {
+  'worklet';
   return rnexecutorchJsi.getExecuTorchRegisteredBackends();
 }
 

From be1942690cd31bda191c1bfee8f01fcedbda51dd Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 17:06:06 +0200
Subject: [PATCH 11/38] docs(cv): standardize TSDoc comments, extract task
 options, and document tensor shapes/dtypes

---
 .../src/extensions/cv/index.ts                |   6 +
 .../src/extensions/cv/ops/boxes.ts            |  92 ++++++------
 .../src/extensions/cv/ops/image.ts            | 140 ++++++++++++------
 .../src/extensions/cv/ops/points.ts           |  32 ++--
 .../src/extensions/cv/tasks/classification.ts |  41 +++--
 .../src/extensions/cv/tasks/imageEmbedding.ts |  15 +-
 .../cv/tasks/instanceSegmentation.ts          |  52 +++++--
 .../extensions/cv/tasks/keypointDetection.ts  |  45 ++++--
 .../extensions/cv/tasks/objectDetection.ts    |  47 ++++--
 .../src/extensions/cv/tasks/preprocessing.ts  |  25 ++--
 .../extensions/cv/tasks/sdxsTextToImage.ts    |  11 ++
 .../cv/tasks/semanticSegmentation.ts          |  18 ++-
 .../src/extensions/cv/tasks/styleTransfer.ts  |  15 +-
 13 files changed, 375 insertions(+), 164 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/index.ts b/packages/react-native-executorch/src/extensions/cv/index.ts
index 58aa2b181d..ac585f06c3 100644
--- a/packages/react-native-executorch/src/extensions/cv/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/index.ts
@@ -1,3 +1,9 @@
+/**
+ * Computer Vision extension providing image buffers, spatial transformations,
+ * OpenCV operators, bounding box utilities, and preprocessing pipelines.
+ * @module CV
+ */
+
 export * from './image';
 export * from './ops';
 export * from './utils/paddleOcrUtils';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index 1151c4a30a..500aa51ea9 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -1,3 +1,9 @@
+/**
+ * Bounding box coordinate decoding, coordinate transforms, and Non-Maximum
+ * Suppression (NMS).
+ * @module CV/Ops/Boxes
+ */
+
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ResizeMode } from './image';
@@ -7,26 +13,11 @@ import { scalePoint } from './points';
  * Mapping of bounding box formats to their coordinate representations.
  * @category Types
  */
-export type BoxMap = {
-  xyxy: {
-    readonly xmin: number;
-    readonly ymin: number;
-    readonly xmax: number;
-    readonly ymax: number;
-  };
-  xywh: {
-    readonly xmin: number;
-    readonly ymin: number;
-    readonly w: number;
-    readonly h: number;
-  };
-  cxcywh: {
-    readonly cx: number;
-    readonly cy: number;
-    readonly w: number;
-    readonly h: number;
-  };
-};
+export type BoxMap = Readonly<{
+  xyxy: Readonly<{ xmin: number; ymin: number; xmax: number; ymax: number }>;
+  xywh: Readonly<{ xmin: number; ymin: number; w: number; h: number }>;
+  cxcywh: Readonly<{ cx: number; cy: number; w: number; h: number }>;
+}>;
 
 /**
  * The formats of bounding boxes.
@@ -39,9 +30,22 @@ export type BoxFormat = keyof BoxMap;
  * @category Types
  */
 export type BoundingBox<F extends BoxFormat> = F extends any
-  ? { readonly format: F } & Readonly<BoxMap[F]>
+  ? { readonly format: F } & BoxMap[F]
   : never;
 
+/**
+ * Configuration options for scaling bounding box coordinates.
+ * @category Types
+ */
+export type ScaleBoxOptions = {
+  /** The source bounds (e.g. model input dimensions). */
+  readonly from: { readonly width: number; readonly height: number };
+  /** The destination bounds (e.g. original image dimensions). */
+  readonly to: { readonly width: number; readonly height: number };
+  /** The mode used to resize the image (excluding `'crop'`). */
+  readonly resizeMode: Exclude<ResizeMode, 'crop'>;
+};
+
 /**
  * Decodes bounding box coordinates from a 4-tuple into a structured BoundingBox
  * object.
@@ -73,19 +77,12 @@ export function decodeBox<F extends BoxFormat>(
  * @typeParam F Bounding box coordinate format.
  * @param box The original BoundingBox.
  * @param options Options defining dimensions and resize modes.
- * @param options.from The source bounds (e.g. model input dimensions).
- * @param options.to The destination bounds (e.g. original image dimensions).
- * @param options.resizeMode The mode used to resize the image {@link ResizeMode}
- * (excluding `'crop'`).
+ * See {@link ScaleBoxOptions}.
  * @returns The scaled BoundingBox object.
  */
 export function scaleBox<F extends BoxFormat>(
   box: BoundingBox<F>,
-  options: {
-    readonly from: { readonly width: number; readonly height: number };
-    readonly to: { readonly width: number; readonly height: number };
-    readonly resizeMode: Exclude<ResizeMode, 'crop'>;
-  }
+  options: ScaleBoxOptions
 ): BoundingBox<F> {
   'worklet';
   const { from, to, resizeMode } = options;
@@ -162,15 +159,12 @@ export type NmsOptions = {
  * Executes Non-Maximum Suppression (NMS) on bounding boxes and confidence
  * scores.
  * @category Utils
- * @param boxes Bounding boxes coordinate tensor.
- * @param scores Bounding boxes confidence scores tensor.
+ * @param boxes Bounding boxes coordinate tensor. Expected shape `[N, 4]` and
+ * data type `float32`.
+ * @param scores Bounding boxes confidence scores tensor. Expected shape `[N]`
+ * (1D) and data type `float32`.
  * @param options Options configuring NMS thresholds and execution mode.
- * @param options.boxFormat The bounding box format {@link BoxFormat}.
- * @param options.iouThreshold Intersection over Union (IoU) threshold for
- * suppression.
- * @param options.confidenceThreshold Minimum confidence score for candidate
- * selection.
- * @param options.nmsType The NMS algorithm variant {@link NmsOptions.nmsType}.
+ * See {@link NmsOptions}.
  * @returns The resulting indices of the non-suppressed boxes:
  * - For `standard` NMS: A 1D array of indices (`number[]`) representing the
  *   selected boxes.
@@ -178,6 +172,9 @@ export type NmsOptions = {
  *   groups of overlapping boxes, where the first element of each group is the
  *   top candidate and the group indices are used to calculate the weighted
  *   average of coordinates.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes or
+ * formats are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function nms(
   boxes: Tensor,
@@ -195,18 +192,23 @@ export function nms(boxes: Tensor, scores: Tensor, options: NmsOptions): number[
 }
 
 /**
- * Masks the source tensor by keeping only the elements inside the specified
- * bounding box, writing the result to a pre-allocated destination tensor.
+ * Masks the source image tensor by keeping only the elements inside the specified
+ * bounding box, writing the result to a pre-allocated destination image tensor.
  *
- * Note: This operation does not change the tensor dimensions (it does not crop
+ * Note: This operation does not change the image tensor dimensions (it does not crop
  * the shape). Instead, it copies the elements within the box coordinates from
  * `src` to `dst`, and sets all elements outside the box to `0`.
  * @category Typescript API
- * @param src The source tensor of shape [H, W, C].
- * @param dst The pre-allocated destination tensor of shape [H, W, C] and the
- * same data type as `src`.
+ * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
+ * (channels-last). Supports any numeric data type.
+ * @param dst The pre-allocated destination image tensor to write masked values to.
+ * Expected shape `[H, W, C]` in HWC layout and the same data type as `src`.
  * @param box The bounding box defining the region of interest to copy.
- * @returns The destination tensor containing the masked output.
+ * @returns The destination image tensor containing the masked output of shape
+ * `[H, W, C]` and matching data type.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * layouts, or data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function restrictToBox(src: Tensor, dst: Tensor, box: BoundingBox<BoxFormat>): Tensor {
   'worklet';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index b1e5a4ee34..77a600486e 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -1,3 +1,12 @@
+/**
+ * Low-level image manipulation and transformation operators.
+ *
+ * Provides native OpenCV-accelerated image operations on tensors, including
+ * spatial resizing, color space conversion, channel transposition (HWC/CHW),
+ * pixel normalization, and colormap application.
+ * @module CV/Ops/Image
+ */
+
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ImageFormat } from '../image';
@@ -74,11 +83,11 @@ export type InterpolationMethod = 'nearest' | 'area' | 'cubic' | 'lanczos' | 'li
  * @category Types
  */
 export type ResizeOptions = {
-  /** How the image is resized {@link ResizeMode}. */
+  /** How the image is resized (stretch, letterbox, or crop). */
   readonly mode?: ResizeMode;
-  /** Background fill value used when letterboxing. */
+  /** Background fill value used when letterboxing (padding). */
   readonly padValue?: number;
-  /** Pixel interpolation method {@link InterpolationMethod}. */
+  /** Pixel interpolation method. */
   readonly interpolation?: InterpolationMethod;
 };
 
@@ -89,10 +98,13 @@ export type ResizeOptions = {
 export type NormalizeOptions = {
   /**
    * Multiplicative coefficient applied as `pixel * alpha`. Single value for
-   * uniform scaling, array for per-channel.
+   * uniform scaling across all channels, or per-channel array.
    */
   readonly alpha?: number | readonly number[];
-  /** Additive offset applied as `pixel * alpha + beta`. Single value or per-channel array. */
+  /**
+   * Additive offset applied as `pixel * alpha + beta`. Single value or
+   * per-channel array.
+   */
   readonly beta?: number | readonly number[];
 };
 
@@ -101,17 +113,21 @@ export type NormalizeOptions = {
  *
  * Supports various {@link ResizeMode} and {@link InterpolationMethod} options.
  * @category Typescript API
- * @param src The source image tensor in HWC layout. Shape [H,W,C].
- * @param dst The pre-allocated destination tensor to write the resized image
- * to. `dst` must be in HWC layout and its number of channels must match `src`.
- * Shape [H',W',C].
- * @param options Configuration options for resizing.
- * @param options.mode The resize algorithm mode {@link ResizeMode}. Defaults to
- * `'stretch'`.
- * @param options.interpolation The pixel interpolation method
- * {@link InterpolationMethod}. Defaults to `'lanczos'`.
- * @param options.padValue Fill value for letterboxing. Defaults to `0`.
- * @returns The destination tensor containing the resized image.
+ * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
+ * (channels-last). Supports any numeric data type (e.g. `uint8`, `float32`).
+ * @param dst The pre-allocated destination image tensor to write the resized
+ * image to. Expected shape `[H', W', C]` in HWC layout (spatial dimensions
+ * `[H', W']`, channel count `C` matching `src`) and the same data type as
+ * `src`.
+ * @param options Configuration options for resizing. When options or any
+ * individual properties are omitted, defaults to `'stretch'` mode, `'lanczos'`
+ * interpolation, and `0` padding.
+ * See {@link ResizeOptions}.
+ * @returns The destination image tensor containing the resized image of shape
+ * `[H', W', C]` and matching data type.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * layouts, or data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function resize(src: Tensor, dst: Tensor, options?: ResizeOptions): Tensor {
   'worklet';
@@ -126,13 +142,19 @@ export function resize(src: Tensor, dst: Tensor, options?: ResizeOptions): Tenso
  * Converts the color space of an image tensor using a specified color
  * conversion code.
  * @category Typescript API
- * @param src The source image tensor in HWC layout. Shape [H,W,C].
- * @param dst The pre-allocated destination tensor to write the converted image
- * to. `dst` must be in HWC layout and its spatial dimensions [H,W] as well as
- * dtype must match `src`. Shape [H,W,C'].
- * @param code The color conversion code indicating source and target spaces
- * (e.g. 'RGBA2RGB').
- * @returns The destination tensor containing the converted image.
+ * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
+ * (channels-last). Supports any numeric data type.
+ * @param dst The pre-allocated destination image tensor to write the converted
+ * image to. Expected shape `[H, W, C']` in HWC layout (spatial dimensions `[H,
+ * W]` and data type matching `src`, with target channel count `C'` determined
+ * by `code`).
+ * @param code The color conversion code indicating source and target spaces.
+ * See {@link ColorConversionCode}.
+ * @returns The destination image tensor containing the converted image of shape
+ * `[H, W, C']` and matching data type.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * layouts, or data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function cvtColor(src: Tensor, dst: Tensor, code: ColorConversionCode): Tensor {
   'worklet';
@@ -145,11 +167,15 @@ export function cvtColor(src: Tensor, dst: Tensor, code: ColorConversionCode): T
  *
  * Commonly required for PyTorch Edge models which expect channels-first inputs.
  * @category Typescript API
- * @param src The source image tensor in HWC layout. Shape [H,W,C].
- * @param dst The pre-allocated destination tensor in CHW layout. `dst` tensor's
- * spatial dimensions [H,W], number of channels and dtype must match `src`.
- * Shape [C,H,W].
- * @returns The destination tensor in CHW layout.
+ * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
+ * (channels-last). Supports any numeric data type.
+ * @param dst The pre-allocated destination image tensor in CHW layout. Expected
+ * shape `[C, H, W]` (channels-first) and the same data type as `src`.
+ * @returns The destination image tensor in CHW layout of shape `[C, H, W]` and
+ * matching data type.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * layouts, or data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function toChannelsFirst(src: Tensor, dst: Tensor): Tensor {
   'worklet';
@@ -163,11 +189,15 @@ export function toChannelsFirst(src: Tensor, dst: Tensor): Tensor {
  * Useful for post-processing model outputs back into channels-last layouts for
  * rendering or display.
  * @category Typescript API
- * @param src The source image tensor in CHW layout. Shape [C,H,W].
- * @param dst The pre-allocated destination tensor in HWC layout. `dst` tensor's
- * spatial dimensions [H,W], number of channels and dtype must match `src`.
- * Shape [H,W,C].
- * @returns The destination tensor in HWC layout.
+ * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
+ * (channels-first). Supports any numeric data type.
+ * @param dst The pre-allocated destination image tensor in HWC layout. Expected
+ * shape `[H, W, C]` (channels-last) and the same data type as `src`.
+ * @returns The destination image tensor in HWC layout of shape `[H, W, C]` and
+ * matching data type.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * layouts, or data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function toChannelsLast(src: Tensor, dst: Tensor): Tensor {
   'worklet';
@@ -181,14 +211,21 @@ export function toChannelsLast(src: Tensor, dst: Tensor): Tensor {
  * uniformly or channel-wise using array options. The result is cast to `dst`
  * tensor's dtype.
  * @category Typescript API
- * @param src The source image tensor in CHW layout. Shape [C,H,W].
- * @param dst The pre-allocated destination tensor to write the normalized
- * values to. `dst` must have the same shape as `src`. Shape [C,H,W].
- * @param options Normalization scaling coefficients.
- * @param options.alpha Multiplicative scaling coefficient(s). Defaults to
- * `1 / 255.0`.
- * @param options.beta Additive offset coefficient(s). Defaults to `0.0`.
- * @returns The destination tensor containing the normalized image.
+ * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
+ * (channels-first). Supports any numeric data type (typically `uint8` or
+ * `float32`).
+ * @param dst The pre-allocated destination image tensor to write normalized
+ * values to. Expected shape `[C, H, W]` matching `src`. The computed values are
+ * cast to `dst` tensor's target data type (typically `float32` or `uint8`).
+ * @param options Normalization scaling coefficients. When options or any
+ * individual properties are omitted, defaults to `alpha: 1 / 255.0` and `beta:
+ * 0.0`.
+ * See {@link NormalizeOptions}.
+ * @returns The destination image tensor containing the normalized image of
+ * shape `[C, H, W]` and target data type.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * layouts, or data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function normalize(src: Tensor, dst: Tensor, options?: NormalizeOptions): Tensor {
   'worklet';
@@ -207,14 +244,19 @@ export function normalize(src: Tensor, dst: Tensor, options?: NormalizeOptions):
  * up its corresponding RGBA color in the provided colormap palette, and writes
  * it to the destination tensor.
  * @category Typescript API
- * @param src The source index/mask tensor. Must be an integer tensor of `int32`
- * dtype containing class indices. Shape `[H, W, 1]` (or `[H, W]`).
- * @param dst The pre-allocated destination tensor to write the mapped RGBA
- * values to. Must be a 3D image tensor in HWC layout and `uint8` dtype. Shape
- * `[H, W, 4]`.
- * @param colormap An array of RGBA color arrays `[R, G, B, A]` corresponding to each
- * class index. The size of this list must cover all class indices present in `src`.
- * @returns The destination tensor with the applied colormap.
+ * @param src The source index/mask image tensor. Expected shape `[H, W, 1]` in
+ * HWC layout with data type `int32` containing class indices.
+ * @param dst The pre-allocated destination image tensor to write the mapped
+ * RGBA values to. Expected shape `[H, W, 4]` in HWC layout with data type
+ * `uint8`.
+ * @param colormap An array of RGBA color arrays `[R, G, B, A]` corresponding to
+ * each class index. The size of this list must cover all class indices present
+ * in `src`.
+ * @returns The destination image tensor with the applied colormap of shape `[H,
+ * W, 4]` and data type `uint8`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * layouts, or data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function applyColormap(
   src: Tensor,
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 4f6218e46c..5685849342 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -1,3 +1,8 @@
+/**
+ * 2D point representation and spatial scaling utilities.
+ * @module CV/Ops/Points
+ */
+
 import type { ResizeMode } from './image';
 
 /**
@@ -35,26 +40,29 @@ export function interpolatePoint(a: Point, b: Point, t: number): Point {
   return { x: a.x + (b.x - a.x) * t, y: a.y + (b.y - a.y) * t };
 }
 
+/**
+ * Configuration options for scaling 2D point coordinates.
+ * @category Types
+ */
+export type ScalePointOptions = {
+  /** The source bounds (e.g. model input dimensions). */
+  readonly from: { readonly width: number; readonly height: number };
+  /** The destination bounds (e.g. original image dimensions). */
+  readonly to: { readonly width: number; readonly height: number };
+  /** The mode used to resize the image (excluding `'crop'`). */
+  readonly resizeMode: Exclude<ResizeMode, 'crop'>;
+};
+
 /**
  * Helper function to scale a 2D point based on resize mode and resolution
  * changes.
  * @category Utils
  * @param point The original coordinate point to scale.
  * @param options Options detailing the scaling factors and resize mode.
- * @param options.from The source bounds (e.g. model input dimensions).
- * @param options.to The destination bounds (e.g. original image dimensions).
- * @param options.resizeMode The mode used to resize the image {@link ResizeMode}
- * (excluding `'crop'`).
+ * See {@link ScalePointOptions}.
  * @returns The scaled coordinate point.
  */
-export function scalePoint(
-  point: Point,
-  options: {
-    readonly from: { readonly width: number; readonly height: number };
-    readonly to: { readonly width: number; readonly height: number };
-    readonly resizeMode: Exclude<ResizeMode, 'crop'>;
-  }
-): Point {
+export function scalePoint(point: Point, options: ScalePointOptions): Point {
   'worklet';
   const { from, to, resizeMode } = options;
   switch (resizeMode) {
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
index 4d5e2b180f..ff84894811 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
@@ -1,3 +1,9 @@
+/**
+ * Image classification task pipeline with integrated preprocessing and softmax
+ * decoding.
+ * @module CV/Tasks/Classification
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -28,13 +34,25 @@ export type ClassifierModel<L> = {
   /** Local path or remote URL of the `.pte` model file. */
   readonly modelPath: string;
   /**
-   * Image preprocessing and label vocabulary
-   * {@link ClassifierOptions}. The `labels` array length must
+   * Image preprocessing and label vocabulary. The `labels` array length must
    * match the model's output dimension.
+   * See {@link ClassifierOptions}.
    */
   readonly modelOpts: ClassifierOptions<L>;
 };
 
+/**
+ * Optional configuration parameters for classification inference.
+ * @category Types
+ */
+export type ClassifyOptions = {
+  /**
+   * Number of top-scoring classification results to return. If omitted, all
+   * classes are returned.
+   */
+  readonly topk?: number;
+};
+
 /**
  * Result structure representing a single classification prediction.
  * @category Types
@@ -57,10 +75,14 @@ export type Classification<L> = {
  * @category Typescript API
  * @typeParam L The type representing the classification labels.
  * @param config Classifier task configuration containing path and options.
+ * See {@link ClassifierModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing classification and
  * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * `SCHEMA_MISMATCH` if model schema does not match classification spec, or
+ * `INVALID_ARGUMENT` if labels length does not match model output classes.
  */
 export async function createClassifier<L>(
   config: ClassifierModel<L>,
@@ -75,18 +97,20 @@ export async function createClassifier<L>(
    * Performs asynchronous image classification on the given input image.
    * @param input The input image buffer.
    * @param options Configuration options for classification.
-   * @param options.topk The number of top-scoring classification results to
-   * return. If omitted, all classes are returned. Must be non-negative.
+   * See {@link ClassifyOptions}.
    * @returns A promise resolving to the list of classifications sorted by
    * confidence.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if `topk` is
+   * negative, `RESOURCE_BUSY` if the model is in use, or
+   * `RESOURCE_DISPOSED` if disposed.
    */
-  classify: (input: ImageBuffer, options?: { topk?: number }) => Promise<Classification<L>[]>;
+  classify: (input: ImageBuffer, options?: ClassifyOptions) => Promise<Classification<L>[]>;
 
   /**
    * Synchronous version of {@link classify} to be executed directly on the
    * caller or worklet thread.
    */
-  classifyWorklet: (input: ImageBuffer, options?: { topk?: number }) => Classification<L>[];
+  classifyWorklet: (input: ImageBuffer, options?: ClassifyOptions) => Classification<L>[];
 }> {
   const { modelPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
@@ -129,10 +153,7 @@ export async function createClassifier<L>(
     model.dispose();
   };
 
-  const classifyWorklet = (
-    input: ImageBuffer,
-    options?: { topk?: number }
-  ): Classification<L>[] => {
+  const classifyWorklet = (input: ImageBuffer, options?: ClassifyOptions): Classification<L>[] => {
     'worklet';
     if (options?.topk !== undefined && options.topk < 0) {
       throw RnExecuTorchError(
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
index cb85e7bdb9..7a1d59d20d 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
@@ -1,3 +1,8 @@
+/**
+ * Image embedding and visual feature extraction task pipeline.
+ * @module CV/Tasks/ImageEmbedding
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -16,8 +21,9 @@ export type ImageEmbedderModel = {
   /** Local path or remote URL of the `.pte` model file. */
   readonly modelPath: string;
   /**
-   * Image preprocessing (resize, color conversion, normalization)
-   * for embedding models {@link ImagePreprocessorOptions}.
+   * Image preprocessing (resize, color conversion, normalization) for embedding
+   * models.
+   * See {@link ImagePreprocessorOptions}.
    */
   readonly modelOpts: ImagePreprocessorOptions;
 };
@@ -33,10 +39,13 @@ export type ImageEmbedderModel = {
  * runs the forward pass, and returns the raw embedding vector.
  * @category Typescript API
  * @param config Image embedder task configuration containing path and options.
+ * See {@link ImageEmbedderModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing the embedding and
  * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match embedding spec.
  */
 export async function createImageEmbedder(
   config: ImageEmbedderModel,
@@ -51,6 +60,8 @@ export async function createImageEmbedder(
    * Asynchronously computes the embedding vector for the given input image.
    * @param input The input image buffer.
    * @returns A promise resolving to the embedding vector.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   embed: (input: ImageBuffer) => Promise<Float32Array>;
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index 68bc1fa00e..fe2d50196b 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -1,3 +1,9 @@
+/**
+ * Instance segmentation task pipeline with NMS, bounding box scaling, and mask
+ * extraction.
+ * @module CV/Tasks/InstanceSegmentation
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -56,13 +62,35 @@ export type InstanceSegmenterModel<F extends BoxFormat, L> = {
   /** Local path or remote URL of the `.pte` model file. */
   readonly modelPath: string;
   /**
-   * Image preprocessing, label vocabulary, bounding box format,
-   * and default NMS/mask/confidence thresholds
-   * {@link InstanceSegmenterOptions}.
+   * Image preprocessing, label vocabulary, bounding box format, and default
+   * NMS/mask/confidence thresholds.
+   * See {@link InstanceSegmenterOptions}.
    */
   readonly modelOpts: InstanceSegmenterOptions<F, L>;
 };
 
+/**
+ * Optional configuration parameters for instance segmentation inference.
+ * @category Types
+ */
+export type SegmentInstancesOptions = {
+  /**
+   * Minimum confidence threshold. If omitted, uses
+   * `modelOpts.defaultConfidenceThreshold`.
+   */
+  readonly confidenceThreshold?: number;
+  /**
+   * Intersection over Union (IoU) threshold in NMS. If omitted, uses
+   * `modelOpts.defaultIouThreshold`.
+   */
+  readonly iouThreshold?: number;
+  /**
+   * Mask binarization probability threshold. If omitted, uses
+   * `modelOpts.defaultMaskThreshold`.
+   */
+  readonly maskThreshold?: number;
+};
+
 /**
  * Result structure representing a single detected instance with its bounding box,
  * segmentation mask, label, and confidence score.
@@ -92,10 +120,14 @@ export type InstanceSegmentationResult<F extends BoxFormat, L> = {
  * @typeParam F The bounding box format type.
  * @typeParam L The label type.
  * @param config Model configuration containing path and options.
+ * See {@link InstanceSegmenterModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing instance segmentation
  * and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match instance segmentation
+ * spec.
  */
 export async function createInstanceSegmenter<F extends BoxFormat, L>(
   config: InstanceSegmenterModel<F, L>,
@@ -110,17 +142,15 @@ export async function createInstanceSegmenter<F extends BoxFormat, L>(
    * Performs asynchronous instance segmentation on the given input image.
    * @param input The input image buffer.
    * @param options Execution override options.
-   * @param options.confidenceThreshold Minimum confidence threshold. If
-   * omitted, uses `modelOpts.defaultConfidenceThreshold`.
-   * @param options.iouThreshold Intersection over Union (IoU) threshold in NMS. If omitted, uses
-   * `modelOpts.defaultIouThreshold`.
-   * @param options.maskThreshold Mask binarization threshold. If omitted,
-   * uses `modelOpts.defaultMaskThreshold`.
+   * See {@link SegmentInstancesOptions}.
    * @returns A promise resolving to a list of detected instances.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if predicted class
+   * index is out of bounds, `RESOURCE_BUSY` if the model is in use, or
+   * `RESOURCE_DISPOSED` if disposed.
    */
   segmentInstances: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number; iouThreshold?: number; maskThreshold?: number }
+    options?: SegmentInstancesOptions
   ) => Promise<InstanceSegmentationResult<F, L>[]>;
 
   /**
@@ -129,7 +159,7 @@ export async function createInstanceSegmenter<F extends BoxFormat, L>(
    */
   segmentInstancesWorklet: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number; iouThreshold?: number; maskThreshold?: number }
+    options?: SegmentInstancesOptions
   ) => InstanceSegmentationResult<F, L>[];
 }> {
   const { modelPath, modelOpts } = config;
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index 66ac81adb9..004fe4f366 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -1,3 +1,9 @@
+/**
+ * Keypoint and pose detection task pipeline with weighted NMS and landmark
+ * scaling.
+ * @module CV/Tasks/KeypointDetection
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor, type Tensor } from '../../../core/tensor';
@@ -22,7 +28,7 @@ export type KeypointDetectorOptions<F extends BoxFormat, L extends PropertyKey>
   ImagePreprocessorOptions,
   'resizeMode'
 > & {
-  /** Resize mode for preprocessing input images {@link ResizeMode} (excluding `'crop'`). */
+  /** Resize mode for preprocessing input images (excluding `'crop'`). */
   readonly resizeMode: Exclude<ResizeMode, 'crop'>;
   /** How bounding box coordinates are interpreted {@link BoxFormat}. */
   readonly boxFormat: F;
@@ -42,13 +48,30 @@ export type KeypointDetectorModel<F extends BoxFormat, L extends PropertyKey> =
   /** Local path or remote URL of the `.pte` model file. */
   readonly modelPath: string;
   /**
-   * Image preprocessing, landmark names, bounding box format,
-   * and default NMS/confidence thresholds
-   * {@link KeypointDetectorOptions}.
+   * Image preprocessing, landmark names, bounding box format, and default
+   * NMS/confidence thresholds.
+   * See {@link KeypointDetectorOptions}.
    */
   readonly modelOpts: KeypointDetectorOptions<F, L>;
 };
 
+/**
+ * Optional configuration parameters for keypoint detection inference.
+ * @category Types
+ */
+export type DetectKeypointsOptions = {
+  /**
+   * Minimum confidence score threshold for detections. If omitted, uses
+   * `modelOpts.defaultConfidenceThreshold`.
+   */
+  readonly confidenceThreshold?: number;
+  /**
+   * Intersection over Union (IoU) threshold for NMS. If omitted, uses
+   * `modelOpts.defaultIouThreshold`.
+   */
+  readonly iouThreshold?: number;
+};
+
 /**
  * Plural landmarks mapped by their names to coordinates and detection
  * confidence.
@@ -155,10 +178,13 @@ function postprocess<F extends BoxFormat, L extends PropertyKey>(
  * @typeParam F The bounding box format.
  * @typeParam L The landmark labels type.
  * @param config Keypoint task configuration containing path and options.
+ * See {@link KeypointDetectorModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing keypoint detection and
  * disposal bindings.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match keypoint detection spec.
  */
 export async function createKeypointDetector<F extends BoxFormat, L extends PropertyKey>(
   config: KeypointDetectorModel<F, L>,
@@ -174,15 +200,14 @@ export async function createKeypointDetector<F extends BoxFormat, L extends Prop
    * input image.
    * @param input The input image buffer.
    * @param options Configuration options for keypoint detection.
-   * @param options.confidenceThreshold Minimum confidence score for a
-   * detection. If omitted, uses `modelOpts.defaultConfidenceThreshold`.
-   * @param options.iouThreshold Intersection over Union (IoU) threshold for
-   * NMS. If omitted, uses `modelOpts.defaultIouThreshold`.
+   * See {@link DetectKeypointsOptions}.
    * @returns A promise resolving to the list of keypoint detections.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   detectKeypoints: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number; iouThreshold?: number }
+    options?: DetectKeypointsOptions
   ) => Promise<KeypointDetection<F, L>[]>;
 
   /**
@@ -191,7 +216,7 @@ export async function createKeypointDetector<F extends BoxFormat, L extends Prop
    */
   detectKeypointsWorklet: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number; iouThreshold?: number }
+    options?: DetectKeypointsOptions
   ) => KeypointDetection<F, L>[];
 }> {
   const { modelPath, modelOpts } = config;
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index 60cff1b272..8136352d8d 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -1,3 +1,9 @@
+/**
+ * Object detection task pipeline with integrated preprocessing, NMS, and box
+ * scaling.
+ * @module CV/Tasks/ObjectDetection
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -22,7 +28,7 @@ export type ObjectDetectorOptions<F extends BoxFormat, L> = Omit<
   ImagePreprocessorOptions,
   'resizeMode'
 > & {
-  /** Resize mode for preprocessing input images {@link ResizeMode} (excluding `'crop'`). */
+  /** Resize mode for preprocessing input images (excluding `'crop'`). */
   readonly resizeMode: Exclude<ResizeMode, 'crop'>;
   /** Array of class labels matching the model's output vocabulary. */
   readonly labels: readonly L[];
@@ -42,13 +48,30 @@ export type ObjectDetectorModel<F extends BoxFormat, L> = {
   /** Local path or remote URL of the `.pte` model file. */
   readonly modelPath: string;
   /**
-   * Image preprocessing, label vocabulary, and default
-   * NMS/confidence thresholds {@link ObjectDetectorOptions}.
-   * Used as fallbacks when per-call overrides are omitted.
+   * Image preprocessing, label vocabulary, and default NMS/confidence
+   * thresholds. Used as fallbacks when per-call overrides are omitted.
+   * See {@link ObjectDetectorOptions}.
    */
   readonly modelOpts: ObjectDetectorOptions<F, L>;
 };
 
+/**
+ * Optional configuration parameters for object detection inference.
+ * @category Types
+ */
+export type DetectObjectsOptions = {
+  /**
+   * Minimum confidence score threshold. If omitted, uses
+   * `modelOpts.defaultConfidenceThreshold`.
+   */
+  readonly confidenceThreshold?: number;
+  /**
+   * Intersection over Union (IoU) threshold for NMS. If omitted, uses
+   * `modelOpts.defaultIouThreshold`.
+   */
+  readonly iouThreshold?: number;
+};
+
 /**
  * Result structure representing a single object detection prediction.
  * @category Types
@@ -73,10 +96,13 @@ export type ObjectDetection<F extends BoxFormat, L> = {
  * @typeParam F The bounding box format.
  * @typeParam L The type representing the class labels.
  * @param config Object detector task configuration containing path and options.
+ * See {@link ObjectDetectorModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing object detection and
  * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match object detection spec.
  */
 export async function createObjectDetector<F extends BoxFormat, L>(
   config: ObjectDetectorModel<F, L>,
@@ -88,17 +114,18 @@ export async function createObjectDetector<F extends BoxFormat, L>(
   dispose: () => void;
 
   /**
+   * Asynchronously performs object detection on the input image.
    * @param input The input image buffer.
    * @param options Configuration options for object detection.
-   * @param options.confidenceThreshold Minimum confidence score threshold. If
-   * omitted, uses `modelOpts.defaultConfidenceThreshold`.
-   * @param options.iouThreshold Intersection over Union (IoU) threshold. If
-   * omitted, uses `modelOpts.defaultIouThreshold`.
+   * See {@link DetectObjectsOptions}.
    * @returns A promise resolving to the list of object detections.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if predicted class
+   * index is out of bounds, `RESOURCE_BUSY` if the model is in use, or
+   * `RESOURCE_DISPOSED` if disposed.
    */
   detectObjects: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number; iouThreshold?: number }
+    options?: DetectObjectsOptions
   ) => Promise<ObjectDetection<F, L>[]>;
 
   /**
@@ -107,7 +134,7 @@ export async function createObjectDetector<F extends BoxFormat, L>(
    */
   detectObjectsWorklet: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number; iouThreshold?: number }
+    options?: DetectObjectsOptions
   ) => ObjectDetection<F, L>[];
 }> {
   const { modelPath, modelOpts } = config;
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
index fcda285da8..b3b7314620 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
@@ -1,3 +1,8 @@
+/**
+ * Reusable image preprocessing pipeline for neural network inputs.
+ * @module CV/Tasks/Preprocessing
+ */
+
 import { tensor, type Tensor } from '../../../core/tensor';
 
 import type { ImageBuffer } from '../image';
@@ -19,16 +24,13 @@ import { RnExecuTorchError } from '../../../core/error';
  * @category Types
  */
 export type ImagePreprocessorOptions = {
-  /**
-   * How the input image is resized to match the model's expected
-   * dimensions {@link ResizeMode}.
-   */
+  /** How the input image is resized to match the model's expected dimensions. */
   readonly resizeMode: ResizeMode;
-  /** Algorithm used when resizing {@link InterpolationMethod}. `'linear'` is a good default. */
+  /** Algorithm used when resizing (e.g. `'linear'`, `'lanczos'`). */
   readonly interpolation: InterpolationMethod;
   /** Normalization scaling coefficients. */
   readonly normalizeOpts: NormalizeOptions;
-  /** Optional background fill value used when letterboxing. */
+  /** Optional background fill value used when letterboxing (padding). */
   readonly padValue?: number;
 };
 
@@ -41,11 +43,14 @@ export type ImagePreprocessorOptions = {
  * disposed of when calling `dispose()`.
  * @category Typescript API
  * @param options Normalization scaling coefficients, interpolation algorithms, and
- * crop/resize modes.
- * @param outputShape Expected output shape of the model input tensor (must
- * match `[1, 3, H, W]` or `[3, H, W]`).
+ * resize modes.
+ * See {@link ImagePreprocessorOptions}.
+ * @param outputShape Expected output shape of the preprocessed model input
+ * tensor (must match rank-3 `[3, H, W]` or rank-4 `[1, 3, H, W]`).
  * @returns An object containing the `process` runner function and a `dispose`
  * method.
+ * @throws {RnExecuTorchError} With code `SCHEMA_MISMATCH` if `outputShape` does
+ * not match rank-3 `[3, H, W]` or rank-4 `[1, 3, H, W]`.
  */
 export function createImagePreprocessor(
   options: ImagePreprocessorOptions,
@@ -60,7 +65,7 @@ export function createImagePreprocessor(
    * need to dispose of it manually.
    * @param input The input image buffer to preprocess.
    * @returns A reference to the output tensor containing preprocessed float32
-   * data.
+   * data of shape `[3, H, W]` (or `[1, 3, H, W]`) and data type `float32`.
    */
   process: (input: ImageBuffer) => Tensor;
   /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
index 0130e6cd07..379696d5ad 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
@@ -1,3 +1,9 @@
+/**
+ * SDXS single-step text-to-image generation pipeline with latent diffusion and
+ * TAESD decoding.
+ * @module CV/Tasks/SdxsTextToImage
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -51,8 +57,11 @@ export type SdxsTextToImageModel = {
  * tensors, and registers disposal hooks that release all native memory.
  * @category Typescript API
  * @param config SDXS pipeline configuration containing the model and tokenizer paths.
+ * See {@link SdxsTextToImageModel}.
  * @param runtime Optional worklet runtime thread on which to run generation.
  * @returns A promise resolving to an object with generation and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model or tokenizer
+ * fails to load, or `SCHEMA_MISMATCH` if model schema does not match SDXS spec.
  */
 export async function createSdxsTextToImage(
   config: SdxsTextToImageModel,
@@ -67,6 +76,8 @@ export async function createSdxsTextToImage(
    * @param seed Seed for the initial latent noise (same seed → same image).
    * Defaults to a time-based value so omitting it yields a fresh image each call.
    * @returns A promise resolving to the generated RGBA image buffer.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   generate: (prompt: string, seed?: number) => Promise<ImageBuffer>;
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
index 85961ff662..1731817c98 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
@@ -1,3 +1,9 @@
+/**
+ * Semantic segmentation task pipeline with colormap mapping and output mask
+ * rendering.
+ * @module CV/Tasks/SemanticSegmentation
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -40,9 +46,9 @@ export type SemanticSegmenterModel<L> = {
   /** Local path or remote URL of the `.pte` model file. */
   readonly modelPath: string;
   /**
-   * Image preprocessing, output mask interpolation, and label
-   * vocabulary {@link SemanticSegmenterOptions}. `resizeMode`
-   * is fixed to `'stretch'`.
+   * Image preprocessing, output mask interpolation, and label vocabulary.
+   * `resizeMode` is fixed to `'stretch'`.
+   * See {@link SemanticSegmenterOptions}.
    */
   readonly modelOpts: SemanticSegmenterOptions<L>;
 };
@@ -84,9 +90,13 @@ function hslToRgb(h: number, s: number, l: number): [number, number, number] {
  * @category Typescript API
  * @typeParam L The type representing the segmentation labels.
  * @param config Segmenter task configuration containing path and options.
+ * See {@link SemanticSegmenterModel}.
  * @param runtime Optional worklet runtime thread environment context.
  * @returns A promise resolving to an object containing segmentation and
  * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * `SCHEMA_MISMATCH` if model schema does not match segmentation spec, or
+ * `INVALID_ARGUMENT` if labels length does not match model output classes.
  */
 export async function createSemanticSegmenter<L extends PropertyKey = string>(
   config: SemanticSegmenterModel<L>,
@@ -117,6 +127,8 @@ export async function createSemanticSegmenter<L extends PropertyKey = string>(
    * provided, any labels omitted from it will default to being rendered as
    * fully transparent.
    * @returns A promise resolving to the segmentation result.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   segment: (
     input: ImageBuffer,
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
index a15d537ec2..80642340c3 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
@@ -1,3 +1,9 @@
+/**
+ * Neural style transfer task pipeline with output rendering and colorspace
+ * conversion.
+ * @module CV/Tasks/StyleTransfer
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -37,9 +43,9 @@ export type StyleTransferModel = {
   /** Local path or remote URL of the `.pte` model file. */
   readonly modelPath: string;
   /**
-   * Input preprocessing and output postprocessing
-   * {@link StyleTransferOptions} (normalization back to uint8,
+   * Input preprocessing and output postprocessing (normalization back to uint8,
    * interpolation). `resizeMode` is fixed to `'stretch'`.
+   * See {@link StyleTransferOptions}.
    */
   readonly modelOpts: StyleTransferOptions;
 };
@@ -52,8 +58,11 @@ export type StyleTransferModel = {
  * registers clean disposal hooks to clear all native memory.
  * @category Typescript API
  * @param config Style transfer task configuration containing path and options.
+ * See {@link StyleTransferModel}.
  * @param runtime Optional worklet runtime thread on which to run the model execution.
  * @returns A promise resolving to an object containing style transfer and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match style transfer spec.
  */
 export async function createStyleTransfer(
   config: StyleTransferModel,
@@ -68,6 +77,8 @@ export async function createStyleTransfer(
    * Performs asynchronous image style transfer on the given input image.
    * @param input The input image buffer.
    * @returns A promise resolving to the styled image buffer.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   transferStyle: (input: ImageBuffer) => Promise<ImageBuffer>;
 

From 65e9adf5a8db87b111e80649193eb6256fd2a12d Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 17:17:29 +0200
Subject: [PATCH 12/38] docs(speech): standardize TSDoc annotations, tensor
 specs, and error throws

---
 .../src/extensions/speech/index.ts            |  6 +++
 .../tasks/fsmnVoiceActivityDetection.ts       | 30 ++++++++++--
 .../speech/tasks/kokoroTextToSpeech.ts        | 18 +++++--
 .../speech/tasks/supertonicTextToSpeech.ts    | 19 +++++---
 .../speech/tasks/whisperSpeechToText.ts       | 35 ++++++++++++--
 .../extensions/speech/utils/kokoroUtils.ts    |  5 ++
 .../src/extensions/speech/utils/phonemizer.ts | 47 ++++++++++++++-----
 .../speech/utils/supertonicUtils.ts           | 10 ++++
 .../speech/utils/textPartitioner.ts           | 25 ++++++----
 .../src/extensions/speech/utils/vadUtils.ts   | 41 +++++++++-------
 10 files changed, 180 insertions(+), 56 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/speech/index.ts b/packages/react-native-executorch/src/extensions/speech/index.ts
index 42c3635917..eacd0f0174 100644
--- a/packages/react-native-executorch/src/extensions/speech/index.ts
+++ b/packages/react-native-executorch/src/extensions/speech/index.ts
@@ -1,3 +1,9 @@
+/**
+ * Speech recognition, text-to-speech synthesis, phonemization, and audio
+ * processing pipelines.
+ * @module Speech
+ */
+
 export * from './utils/vadUtils';
 export * from './utils/supertonicUtils';
 export * from './utils/kokoroUtils';
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
index 3f5ec9cf3a..97cb860246 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
@@ -1,9 +1,13 @@
+/**
+ * Voice Activity Detection (VAD) task pipeline using the FSMN-VAD model.
+ * @module Speech/Tasks/FsmnVoiceActivityDetection
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
 import { loadModel } from '../../../core/model';
 import { validateSpec, DynamicDim as Dyn, method, f32 } from '../../../core/schema';
-
 import { wrapAsync } from '../../../core/runtime';
 import { extractFrames } from '../utils/vadUtils';
 
@@ -177,11 +181,15 @@ function postprocess(scores: Float32Array, options: Required<VadOptions>): Segme
  * segment postprocessing — runs in TypeScript on top of the core `model.execute`
  * primitive.
  * @category Typescript API
- * @param config VAD task configuration containing the model path.
+ * @param config VAD task configuration containing the model path and default
+ * options. See {@link FsmnVadModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing detection and disposal
  * controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model fails to
+ * load, or `SCHEMA_MISMATCH` if the model schema does not match the VAD
+ * specification.
  */
 export async function createFsmnVoiceActivityDetector(
   config: FsmnVadModel,
@@ -195,15 +203,24 @@ export async function createFsmnVoiceActivityDetector(
   /**
    * Asynchronously detects speech segments within a mono waveform sampled at
    * {@link FSMN_VAD_SAMPLE_RATE_HZ}.
-   * @param waveform The input audio samples.
+   * @param waveform The input audio samples (16 kHz mono Float32 PCM).
    * @param options Optional per-call overrides of the detection thresholds.
+   * See {@link VadOptions}.
    * @returns A promise resolving to the detected speech segments, in seconds.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   detectVoice: (waveform: Float32Array, options?: VadOptions) => Promise<Segment[]>;
 
   /**
    * Synchronous version of {@link detectVoice} to be executed directly on the caller
    * or worklet thread.
+   * @param waveform The input audio samples (16 kHz mono Float32 PCM).
+   * @param options Optional per-call overrides of the detection thresholds.
+   * See {@link VadOptions}.
+   * @returns The detected speech segments, in seconds.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   detectVoiceWorklet: (waveform: Float32Array, options?: VadOptions) => Segment[];
 
@@ -212,8 +229,13 @@ export async function createFsmnVoiceActivityDetector(
    * that window and reports a {@link VadEvent} when speech starts or stops,
    * otherwise `undefined`. Designed to be driven straight from a recorder
    * callback. Detection runs synchronously on the calling thread (a few ms).
-   * @param chunk The newly captured audio samples.
+   * @param chunk The newly captured audio samples (16 kHz mono Float32 PCM).
    * @param options Optional overrides of the detection thresholds and margin.
+   * See {@link VadStreamOptions}.
+   * @returns `'speechStart'` or `'speechEnd'` if state transitioned, otherwise
+   * `undefined`.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   detectVoiceOnStream: (chunk: Float32Array, options?: VadStreamOptions) => VadEvent | undefined;
 
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
index ee3d0459d7..60882c3085 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
@@ -1,3 +1,8 @@
+/**
+ * Kokoro Text-to-Speech (TTS) synthesis task pipeline.
+ * @module Speech/Tasks/KokoroTextToSpeech
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import RNBlobUtil from 'react-native-blob-util';
@@ -110,8 +115,12 @@ export type KokoroTtsChunk = {
  * @category Typescript API
  * @typeParam K Voice keys record constraint.
  * @param config Kokoro TTS pipeline configuration containing model and asset paths.
+ * See {@link KokoroTtsModel}.
  * @param runtime Optional worklet runtime thread on which to run inference.
  * @returns A promise resolving to an object with audio synthesis and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if models or voice files
+ * fail to load, `SCHEMA_MISMATCH` if model schemas do not match the Kokoro
+ * specification, or `INVALID_STATE` if phonemizer native support is missing.
  */
 export async function createKokoroTextToSpeech<K extends PropertyKey>(
   config: KokoroTtsModel<K>,
@@ -123,12 +132,11 @@ export async function createKokoroTextToSpeech<K extends PropertyKey>(
   /**
    * Streams synthesized audio chunks as an async generator as each text chunk finishes.
    * @param text Input text (or IPA phonemes) to synthesize into speech.
-   * @param options Per-call execution options.
-   * @param options.voice Voice name, one of the keys of the config's `voices`.
-   * @param options.speed Speech speed factor (range: 0.1 to 3.0). Defaults to 1.0.
-   * @param options.phonemize Whether to phonemize the input. Defaults to true.
-   * @param options.maxChunkLength Maximum phoneme count per chunk.
+   * @param options Per-call execution options. See {@link KokoroTtsOptions}.
    * @returns An AsyncGenerator yielding {@link KokoroTtsChunk} audio buffers.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the voice is unknown
+   * or speed is out of range, `RESOURCE_BUSY` if the model is in use, or
+   * `RESOURCE_DISPOSED` if disposed.
    */
   synthesize: (text: string, options: KokoroTtsOptions<K>) => AsyncGenerator<KokoroTtsChunk>;
 
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
index 68cacbe917..9bfe32f3a1 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
@@ -1,3 +1,8 @@
+/**
+ * Supertonic Text-to-Speech (TTS) synthesis task pipeline.
+ * @module Speech/Tasks/SupertonicTextToSpeech
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import RNBlobUtil from 'react-native-blob-util';
@@ -127,8 +132,12 @@ export type SupertonicTtsChunk = {
  * @category Typescript API
  * @typeParam K Voice style keys record constraint.
  * @param config Supertonic TTS pipeline configuration containing model and asset paths.
+ * See {@link SupertonicTtsModel}.
  * @param runtime Optional worklet runtime thread on which to run inference.
  * @returns A promise resolving to an object with audio synthesis and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if models, indexer, or
+ * voice styles fail to load, or `SCHEMA_MISMATCH` if model schemas do not match
+ * the Supertonic specification.
  */
 export async function createSupertonicTextToSpeech<K extends PropertyKey>(
   config: SupertonicTtsModel<K>,
@@ -140,13 +149,11 @@ export async function createSupertonicTextToSpeech<K extends PropertyKey>(
   /**
    * Streams synthesized audio chunks as an async generator as each text chunk finishes.
    * @param text Input text string to synthesize into speech.
-   * @param options Per-call execution options.
-   * @param options.voiceStyle Voice style key or {@link SupertonicVoiceStyle} object.
-   * @param options.speed Speech speed factor (range: 0.7 to 2.0). Defaults to 1.05.
-   * @param options.totalSteps Number of flow-matching denoising steps. Defaults to 8.
-   * @param options.lang Language ISO code. Defaults to 'na'.
-   * @param options.maxChunkLength Maximum character limit per text chunk. Defaults to 300 (120 for 'ko').
+   * @param options Per-call execution options. See {@link SupertonicTtsOptions}.
    * @returns An AsyncGenerator yielding {@link SupertonicTtsChunk} audio buffers.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if voice style is
+   * invalid or language is unsupported, `RESOURCE_BUSY` if the model is in use,
+   * or `RESOURCE_DISPOSED` if disposed.
    */
   synthesize: (
     text: string,
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
index 01ed991102..837190bc78 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
@@ -1,3 +1,8 @@
+/**
+ * Whisper Speech-to-Text (STT) transcription and live-streaming task pipeline.
+ * @module Speech/Tasks/WhisperSpeechToText
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 import { scheduleOnRN, createSynchronizable } from 'react-native-worklets';
 
@@ -98,12 +103,17 @@ export type WhisperSttModel<L extends WhisperLanguage = WhisperLanguage> = {
 /**
  * Loads a Whisper model and returns a set of transcription helpers.
  * @category Typescript API
- * @param config Model paths and supported-language metadata.
+ * @typeParam L Supported language codes constraint.
+ * @param config Model paths, supported-language metadata, and VAD
+ * configuration.
  * See {@link WhisperSttModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing transcription and
  * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model, tokenizer,
+ * or VAD fails to load, or `SCHEMA_MISMATCH` if model schemas do not match the
+ * Whisper specification.
  */
 export async function createWhisperSpeechToText<L extends WhisperLanguage = WhisperLanguage>(
   config: WhisperSttModel<L>,
@@ -117,11 +127,14 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
   /**
    * Asynchronously transcribes a pre-recorded mono waveform sampled at
    * {@link WHISPER_SAMPLE_RATE_HZ}.
-   * @param audio Raw 16 kHz mono PCM samples.
-   * @param options Transcription options.
+   * @param audio Raw 16 kHz mono PCM audio samples (Float32Array).
+   * @param options Transcription options. See {@link WhisperSttOptions}.
    * @param onToken Optional callback fired on the RN thread for each decoded
    * token.
    * @returns A promise resolving to the full transcript string.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the language is
+   * unsupported, `RESOURCE_BUSY` if the model is in use, or `RESOURCE_DISPOSED`
+   * if disposed.
    */
   transcribe: (
     audio: Float32Array,
@@ -132,6 +145,14 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
   /**
    * Synchronous version of {@link transcribe} to be executed directly on the
    * caller or worklet thread.
+   * @param audio Raw 16 kHz mono PCM audio samples (Float32Array).
+   * @param options Transcription options. See {@link WhisperSttOptions}.
+   * @param onToken Optional callback fired on the RN thread for each decoded
+   * token.
+   * @returns The full transcript string.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the language is
+   * unsupported, `RESOURCE_BUSY` if the model is in use, or `RESOURCE_DISPOSED`
+   * if disposed.
    */
   transcribeWorklet: (
     audio: Float32Array,
@@ -151,6 +172,11 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
    * finalized transcript so far; `nonCommitted` is the in-progress text that
    * may still change.
    * @param options Stream options (language and optional VAD tuning).
+   * See {@link WhisperStreamOptions}.
+   * @returns An AsyncGenerator yielding transcript updates.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the language is
+   * unsupported, `RESOURCE_BUSY` if the model is in use, or `RESOURCE_DISPOSED`
+   * if disposed.
    */
   stream: (
     options: WhisperStreamOptions<L>
@@ -165,7 +191,8 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
   /**
    * Appends a new PCM chunk to the live streaming buffer consumed by
    * {@link stream}. Ignored when streaming is not active.
-   * @param audioChunk The newly captured audio samples.
+   * @param audioChunk The newly captured audio samples (16 kHz mono Float32
+   * PCM).
    */
   streamInsert: (audioChunk: Float32Array) => void;
 }> {
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
index 36d9896bc1..42d964d7c2 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
@@ -1,3 +1,8 @@
+/**
+ * Utility functions, vocoder constants, and voice decoders for Kokoro TTS.
+ * @module Speech/Utils/Kokoro
+ */
+
 /**
  * Number of audio samples generated per single predicted duration tick.
  * @category Constants
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts b/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
index e732ad3c89..8a3cc2f465 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
@@ -1,32 +1,52 @@
+/**
+ * Grapheme-to-phoneme (G2P) conversion using native Phonemis bindings.
+ * @module Speech/Utils/Phonemizer
+ */
+
 import { rnexecutorchJsi } from '../../../native/bridge';
 import { RnExecuTorchError } from '../../../core/error';
 
 declare const phonemizerBrand: unique symbol;
 
 /**
- * Union of all (currently) supported languages in our G2P pipeline.
+ * Union of all supported language codes in the G2P pipeline.
+ * @category Types
  */
 export type PhonemizerLanguage = 'en-us' | 'en-gb' | 'fr' | 'es' | 'it' | 'pt' | 'de' | 'pl' | 'hi';
 
 /**
- * A configuration type compatible with the underlying
- * Phonemis library interface.
+ * Configuration options and asset paths for initializing a {@link Phonemizer}.
+ * @category Types
  */
 export type PhonemizerConfig = {
-  lang: PhonemizerLanguage;
-  taggerSource?: string;
-  lexiconSource?: string;
-  neuralModelSource?: string;
+  /** Target language code to configure the G2P rules for. */
+  readonly lang: PhonemizerLanguage;
+  /** Optional local file path to the part-of-speech tagger model data. */
+  readonly taggerSource?: string;
+  /** Optional local file path to the pronunciation lexicon dictionary. */
+  readonly lexiconSource?: string;
+  /** Optional local file path to the neural G2P model data. */
+  readonly neuralModelSource?: string;
 };
 
+/**
+ * Native Grapheme-to-Phoneme (G2P) conversion interface.
+ * @category Types
+ */
 export type Phonemizer = {
   /**
-   * A standard G2P (grapheme to phoneme) utility.
-   * @param text Input text to be phonemized.
+   * Converts input text into phonetic IPA transcription.
+   * @param text Input text string to be phonemized.
+   * @returns Phonetic transcription string.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the phonemizer is
+   * in use, or `RESOURCE_DISPOSED` if disposed.
    */
   phonemize(text: string): string;
 
-  /** Releases the native phonemizer. The instance must not be used afterwards. */
+  /**
+   * Releases the allocated native phonemizer resources. The instance must not
+   * be used afterwards.
+   */
   dispose(): void;
 
   /**
@@ -37,10 +57,13 @@ export type Phonemizer = {
 };
 
 /**
- * Creates a grapheme-to-phoneme pipeline for the configured language.
+ * Creates a grapheme-to-phoneme (G2P) pipeline for the configured language.
  * @category Typescript API
- * @param config The phonemizer configuration and asset paths.
+ * @param config Phonemizer configuration and asset paths.
+ * See {@link PhonemizerConfig}.
  * @returns The native {@link Phonemizer} instance.
+ * @throws {RnExecuTorchError} With code `INVALID_STATE` if the native build
+ * lacks phonemizer support, or `LOAD_FAILED` if phonemizer assets fail to load.
  */
 export function createPhonemizer(config: PhonemizerConfig): Phonemizer {
   'worklet';
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
index a2fd48745a..fb99a2aa8a 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
@@ -1,3 +1,9 @@
+/**
+ * Text normalization, chunk formatting, and voice style parsing utilities for
+ * Supertonic TTS.
+ * @module Speech/Utils/Supertonic
+ */
+
 import { RnExecuTorchError } from '../../../core/error';
 /**
  * Ported from supertone-inc/supertonic (MIT License)
@@ -116,6 +122,8 @@ export function cleanText(text: string): string {
  * @param chunk The partitioned text chunk.
  * @param lang The language code.
  * @returns The formatted chunk ready for model input.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if `lang` is
+ * unsupported.
  */
 export function formatChunk(chunk: string, lang?: string): string {
   'worklet';
@@ -183,6 +191,8 @@ export type SupertonicVoiceStyle = {
  * @category Utils
  * @param json The parsed JSON voice style object.
  * @returns Parsed SupertonicVoiceStyle containing styleTtl and styleDp Float32Arrays.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the voice style JSON
+ * format is invalid or missing required tensor data.
  */
 export function parseVoiceStyle(json: any): SupertonicVoiceStyle {
   'worklet';
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts b/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
index a047c35b01..1f6844f99f 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
@@ -1,3 +1,9 @@
+/**
+ * Dynamic programming text partitioning algorithm for splitting long sentences
+ * into natural speech chunks.
+ * @module Speech/Utils/TextPartitioner
+ */
+
 import { RnExecuTorchError } from '../../../core/error';
 type Tag = 'eos' | 'pause' | 'whitespace';
 
@@ -28,21 +34,19 @@ export type PartitionOptions = {
   /**
    * Whether to prioritize shorter initial segment lengths and scale up
    * progressively to minimize Time To First Audio (TTFA).
-   * @default false
    */
   readonly prioritizeInitialTtfa?: boolean;
 
   /**
    * Scaling multiplier for the length deviation penalty when the first segment
    * is shorter than target length. Lower values reduce the length penalty for
-   * short initial chunks when prioritizeInitialTtfa is true.
-   * @default 0.1
+   * short initial chunks when `prioritizeInitialTtfa` is true.
    */
   readonly initialShortDeviationScale?: number;
 
   /**
-   * Custom separator penalties for breakpoint tag types ('eos', 'pause',
-   * 'whitespace'). Default: { eos: 5, pause: 80, whitespace: 1000 }.
+   * Custom separator penalties for breakpoint tag types (`'eos'`, `'pause'`,
+   * `'whitespace'`).
    */
   readonly separatorPenalties?: {
     readonly eos?: number;
@@ -87,13 +91,16 @@ function sliceAtCuts(text: string, cutIndices: number[]): string[] {
 }
 
 /**
- * Divides input text into logical segments under the maximum limit using a
- * forward dynamic programming algorithm.
+ * Divides input text into logical segments under the maximum character limit
+ * using a forward dynamic programming algorithm.
  * @category Utils
- * @param text The input text to partition.
- * @param limit The character limit per partition.
+ * @param text The input text string to partition.
+ * @param limit The maximum character limit per partition (must be >= 10).
  * @param options Optional configuration for TTFA prioritization and custom penalties.
+ * See {@link PartitionOptions}.
  * @returns An array of partitioned text segments.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if `limit` is below
+ * 10 or the text cannot be divided into chunks within the given limit.
  */
 export function partition(text: string, limit: number, options?: PartitionOptions): string[] {
   if (!text) {
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
index f89207f4e3..e563af89f0 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
@@ -1,3 +1,8 @@
+/**
+ * Audio framing and pre-emphasis feature extraction for Voice Activity Detection.
+ * @module Speech/Utils/VAD
+ */
+
 import { rnexecutorchJsi } from '../../../native/bridge';
 import { type Tensor } from '../../../core/tensor';
 
@@ -15,23 +20,27 @@ export type ExtractFramesOptions = {
 };
 
 /**
- * Slices a mono `waveform` into `numFrames` overlapping
- * frames, applying per-frame mean-removal, a pre-emphasis filter and the `hann`
- * window, and writing each frame into a zero-padded row of `dst` (shape
- * `[frames, fftLength]`). `dst` is fully zeroed first, so rows beyond
- * `numFrames` stay zero (padding). The frame length is taken from `hann`'s
- * length and the padded width from `dst`'s last dimension.
+ * Slices a mono audio waveform tensor into overlapping frames, applying
+ * per-frame mean-removal, a pre-emphasis filter and a Hann window, and writing
+ * each frame into a zero-padded row of `dst`.
+ *
+ * `dst` is fully zeroed first, so rows beyond `numFrames` stay zero (padding).
+ * The frame length is taken from `hann`'s length and the padded width from
+ * `dst`'s last dimension.
  * @category Typescript API
- * @param waveform Input audio samples, shape `[length]`. Framing starts at the
- * first sample, so pass only the slice to be framed.
- * @param hann Precomputed Hann window, shape `[frameLength]`.
- * @param dst Pre-allocated destination, shape `[frames, fftLength]`.
- * @param options Framing options.
- * @param options.numFrames Number of frames to write (must not exceed `dst`
- * tensor's first dimension `dst.shape[0]`).
- * @param options.hopLength Number of audio samples between consecutive frames.
- * @param options.preemphasis Pre-emphasis filter coefficient.
- * @returns The `dst` tensor, for convenience.
+ * @param waveform Input audio samples tensor. Expected 1D shape `[length]` with
+ * data type `float32`. Framing starts at the first sample.
+ * @param hann Precomputed Hann window tensor. Expected 1D shape `[frameLength]`
+ * with data type `float32`.
+ * @param dst Pre-allocated destination tensor. Expected 2D shape `[frames,
+ * fftLength]` with data type `float32`.
+ * @param options Framing options controlling frame count, hop length, and
+ * pre-emphasis filtering. See {@link ExtractFramesOptions}.
+ * @returns The destination tensor `dst` containing extracted frames of shape
+ * `[frames, fftLength]` and data type `float32`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * data types, or frame windows are invalid, `RESOURCE_BUSY` if a tensor is in
+ * use, or `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function extractFrames(
   waveform: Tensor,

From 179df3d708a49eb5b977180c8a7be0a27b74d0c7 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 17:44:55 +0200
Subject: [PATCH 13/38] docs(math): standardize TSDoc annotations, tensor
 specs, and use random default seed

---
 .../src/extensions/math.ts                    | 103 ++++++++++++------
 1 file changed, 71 insertions(+), 32 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/math.ts b/packages/react-native-executorch/src/extensions/math.ts
index a9f15764c7..cd27320a19 100644
--- a/packages/react-native-executorch/src/extensions/math.ts
+++ b/packages/react-native-executorch/src/extensions/math.ts
@@ -1,15 +1,41 @@
+/**
+ * Native C++ and JavaScript tensor math and random number generation utilities.
+ * @module Math
+ */
+
 import { rnexecutorchJsi } from '../native/bridge';
 import type { Tensor } from '../core/tensor';
 import { RnExecuTorchError } from '../core/error';
 
+/**
+ * Configuration options for normal distribution random number generation.
+ * @category Types
+ */
+export type RandomNormalOptions = {
+  /** The mean of the distribution. */
+  readonly mean?: number;
+  /** The standard deviation of the distribution. */
+  readonly std?: number;
+  /**
+   * The seed for the underlying generator. When omitted, a random seed is
+   * generated so different values are produced each call.
+   */
+  readonly seed?: number;
+};
+
 /**
  * Computes the element-wise sigmoid activation on a float32 source tensor and
  * writes the result to a destination tensor.
  * @category Typescript API
- * @param src The input float32 source tensor. Shape [d1,...,dn].
- * @param dst The pre-allocated float32 destination tensor to write the result
- * to. `dst` tensor must have the same shape as `src`. Shape [d1,...,dn].
- * @returns The destination tensor containing the sigmoid output.
+ * @param src The input source tensor. Expected shape `[d1, ..., dn]` with data
+ * type `float32`.
+ * @param dst The pre-allocated destination tensor to write the result to.
+ * Expected shape `[d1, ..., dn]` matching `src` with data type `float32`.
+ * @returns The destination tensor `dst` containing the sigmoid output of shape
+ * `[d1, ..., dn]` and data type `float32`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes or
+ * data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function sigmoid(src: Tensor, dst: Tensor): Tensor {
   'worklet';
@@ -20,12 +46,17 @@ export function sigmoid(src: Tensor, dst: Tensor): Tensor {
  * Computes the softmax activation along a specified axis on a float32 source
  * tensor and writes the result to a destination tensor.
  * @category Typescript API
- * @param src The input float32 source tensor. Shape [d1,...,dn].
- * @param dst The pre-allocated float32 destination tensor to write the result
- * to. `dst` tensor must have the same shape as `src`. Shape [d1,...,dn].
- * @param axis The dimension along which softmax is computed. Defaults to -1
- * (last dimension).
- * @returns The destination tensor containing the softmax output.
+ * @param src The input source tensor. Expected shape `[d1, ..., dn]` with data
+ * type `float32`.
+ * @param dst The pre-allocated destination tensor to write the result to.
+ * Expected shape `[d1, ..., dn]` matching `src` with data type `float32`.
+ * @param axis The dimension along which softmax is computed. Negative indexing
+ * is supported (e.g. `-1` for the last dimension). Defaults to `-1`.
+ * @returns The destination tensor `dst` containing the softmax output of shape
+ * `[d1, ..., dn]` and data type `float32`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * data types, or `axis` are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function softmax(src: Tensor, dst: Tensor, axis: number = -1): Tensor {
   'worklet';
@@ -36,12 +67,18 @@ export function softmax(src: Tensor, dst: Tensor, axis: number = -1): Tensor {
  * Computes the indices of the maximum values along a specified axis on a
  * float32 source tensor and writes the result to an int32 destination tensor.
  * @category Typescript API
- * @param src The input float32 source tensor. Shape [d1,...,dk,...,dn].
- * @param dst The pre-allocated int32 destination tensor to write the indices
- * to. Shape [d1,...,1,...,dn].
- * @param axis The dimension along which argmax is computed. Defaults to -1
- * (last dimension).
- * @returns The destination tensor containing the argmax output.
+ * @param src The input source tensor. Expected shape `[d1, ..., dk, ..., dn]`
+ * with data type `float32`.
+ * @param dst The pre-allocated destination tensor to write the indices to.
+ * Expected shape `[d1, ..., 1, ..., dn]` (same rank as `src` with dimension 1
+ * along `axis`) and data type `int32`.
+ * @param axis The dimension along which argmax is computed. Negative indexing
+ * is supported (e.g. `-1` for the last dimension). Defaults to `-1`.
+ * @returns The destination tensor `dst` containing the argmax indices of shape
+ * `[d1, ..., 1, ..., dn]` and data type `int32`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * data types, or `axis` are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function argmax(src: Tensor, dst: Tensor, axis: number = -1): Tensor {
   'worklet';
@@ -74,11 +111,17 @@ export function gather(src: Tensor, indices: Tensor, dst: Tensor, axis: number =
  * Applies the element-wise threshold step function on a float32 source tensor and
  * writes the result to a destination tensor.
  * @category Typescript API
- * @param src The input float32 source tensor. Shape [d1,...,dn].
+ * @param src The input source tensor. Expected shape `[d1, ..., dn]` with data
+ * type `float32`.
  * @param dst The pre-allocated destination tensor to write the result to.
- * `dst` tensor must have the same shape as `src` and have dtype float32.
- * @param thresholdVal The threshold value above or equal to which elements are mapped to 1.0.
- * @returns The destination tensor containing the threshold step output.
+ * Expected shape `[d1, ..., dn]` matching `src` with data type `float32`.
+ * @param thresholdVal The threshold value above or equal to which elements are
+ * mapped to 1.0.
+ * @returns The destination tensor `dst` containing the threshold step output of
+ * shape `[d1, ..., dn]` and data type `float32`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes or
+ * data types are invalid, `RESOURCE_BUSY` if a tensor is in use, or
+ * `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function threshold(src: Tensor, dst: Tensor, thresholdVal: number): Tensor {
   'worklet';
@@ -90,8 +133,8 @@ export function threshold(src: Tensor, dst: Tensor, thresholdVal: number): Tenso
  * `[0, 1)`. Unlike `Math.random` it accepts a seed, so a fixed seed yields a
  * reproducible sequence.
  * @category Typescript API
- * @param seed The 32-bit seed for the generator.
- * @returns A function returning the next uniform value in `[0, 1)`.
+ * @param seed The 32-bit integer seed for the generator.
+ * @returns A function returning the next uniform pseudo-random number in `[0, 1)`.
  */
 export function mulberry32(seed: number): () => number {
   'worklet';
@@ -114,21 +157,17 @@ export function mulberry32(seed: number): () => number {
  * {@link mulberry32} so a fixed `seed` reproduces the same sequence.
  * @category Typescript API
  * @param size The number of values to draw.
- * @param options Distribution parameters.
- * @param options.mean The mean of the distribution. Defaults to 0.
- * @param options.std The standard deviation of the distribution. Defaults to 1.
- * @param options.seed The seed for the underlying generator. Defaults to a
- * time-based value, so omitting it produces different values each call.
+ * @param options Distribution parameters. When options or any individual
+ * properties are omitted, defaults to `mean: 0`, `std: 1`, and a random seed.
+ * See {@link RandomNormalOptions}.
  * @returns A `Float32Array` of `size` normally distributed values.
  */
-export function randomNormal(
-  size: number,
-  options?: { mean?: number; std?: number; seed?: number }
-): Float32Array {
+export function randomNormal(size: number, options?: RandomNormalOptions): Float32Array {
   'worklet';
   const mean = options?.mean ?? 0;
   const std = options?.std ?? 1;
-  const uniform = mulberry32(options?.seed ?? Date.now());
+  const seed = options?.seed ?? Math.floor(Math.random() * 0x100000000);
+  const uniform = mulberry32(seed);
   const out = new Float32Array(size);
   for (let i = 0; i < size; i += 2) {
     // Guard against log(0) when the generator returns exactly 0.

From 4a85ef615e190d4e1c1d2f29eb22e28a137d7e62 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 17:59:56 +0200
Subject: [PATCH 14/38] docs(nlp): standardize TSDoc annotations, error throws,
 and simplify worklet comments

---
 .../src/extensions/nlp/index.ts               |  6 +++
 .../src/extensions/nlp/tasks/privacyFilter.ts | 22 +++++---
 .../src/extensions/nlp/tasks/textEmbedding.ts | 13 ++++-
 .../src/extensions/nlp/tasks/tokenization.ts  |  5 ++
 .../src/extensions/nlp/tokenizer.ts           | 51 +++++++++++++++----
 .../nlp/utils/privacyFilterUtils.ts           |  5 ++
 .../tasks/fsmnVoiceActivityDetection.ts       |  6 ---
 .../speech/tasks/whisperSpeechToText.ts       |  8 ---
 8 files changed, 83 insertions(+), 33 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/nlp/index.ts b/packages/react-native-executorch/src/extensions/nlp/index.ts
index b210cb0086..e25a66523f 100644
--- a/packages/react-native-executorch/src/extensions/nlp/index.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/index.ts
@@ -1,2 +1,8 @@
+/**
+ * Natural Language Processing (NLP) tokenization, text embedding, and privacy
+ * filtering pipelines.
+ * @module NLP
+ */
+
 export * from './tokenizer';
 export * from './utils/privacyFilterUtils';
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
index 895600224a..aeaecbcccd 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
@@ -1,3 +1,9 @@
+/**
+ * Privacy Filter task pipeline for detecting personally identifiable
+ * information (PII).
+ * @module NLP/Tasks/PrivacyFilter
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -80,16 +86,17 @@ export type { PiiEntity, PiiEntityType, PiiSegment, ViterbiBiases };
  * edges are discarded in favor of the neighboring window's more centered
  * context.
  * @category Typescript API
+ * @typeParam Label The model's BIOES label space.
  * @param config Privacy filter task configuration containing the model and
- * tokenizer paths plus the label space options.
+ * tokenizer paths plus the label space options. See {@link PrivacyFilterModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing detection and disposal
  * controls.
  * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if `labelNames` is
- * empty or does not start with `'O'`.
- * @throws {RnExecuTorchError} With code `SCHEMA_MISMATCH` if the loaded model's
- * `forward` window is smaller than 2 tokens.
+ * empty or does not start with `'O'`, `LOAD_FAILED` if the model or tokenizer
+ * fails to load, or `SCHEMA_MISMATCH` if the loaded model schema does not match
+ * the privacy filter specification.
  */
 export async function createPrivacyFilter<Label extends string>(
   config: PrivacyFilterModel<Label>,
@@ -102,14 +109,15 @@ export async function createPrivacyFilter<Label extends string>(
 
   /**
    * Asynchronously detects PII entity spans in the given text.
-   * @param input The text to scan for PII.
+   * @param input The text string to scan for PII.
    * @returns A promise resolving to the detected entity spans, in order.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   detectPii: (input: string) => Promise<PiiEntity<PiiEntityType<Label>>[]>;
 
   /**
-   * Synchronous version of {@link detectPii} to be executed directly on the
-   * caller or worklet thread.
+   * Synchronous version of {@link detectPii}.
    */
   detectPiiWorklet: (input: string) => PiiEntity<PiiEntityType<Label>>[];
 }> {
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
index d84774515a..5ede1a72cd 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
@@ -1,3 +1,8 @@
+/**
+ * Text Embedding task pipeline for sentence transformers.
+ * @module NLP/Tasks/TextEmbedding
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
@@ -34,11 +39,14 @@ export type TextEmbedderModel = {
  * pass and returns the raw embedding vector.
  * @category Typescript API
  * @param config Text embedder task configuration containing the model and
- * tokenizer paths.
+ * tokenizer paths. See {@link TextEmbedderModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
  * @returns A promise resolving to an object containing the embedding and
  * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model or tokenizer
+ * fails to load, or `SCHEMA_MISMATCH` if the model schema does not match the
+ * text embedding specification.
  */
 export async function createTextEmbedder(
   config: TextEmbedderModel,
@@ -56,6 +64,9 @@ export async function createTextEmbedder(
    * @param prompt Optional prompt prefix overriding the model's configured
    * `defaultPrompt` for this call.
    * @returns A promise resolving to the embedding vector.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the input text
+   * tokenizes to zero tokens, `RESOURCE_BUSY` if the model is in use, or
+   * `RESOURCE_DISPOSED` if disposed.
    */
   embed: (input: string, prompt?: string) => Promise<Float32Array>;
 
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
index 297028f320..bda219d259 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
@@ -1,3 +1,8 @@
+/**
+ * Tokenization task pipeline wrapping native HuggingFace tokenizers.
+ * @module NLP/Tasks/Tokenization
+ */
+
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { wrapAsync } from '../../../core/runtime';
diff --git a/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts b/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
index abacf62060..c4d4d5bacd 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
@@ -1,3 +1,8 @@
+/**
+ * Native HuggingFace fast tokenizer bindings.
+ * @module NLP/Tokenizer
+ */
+
 import { rnexecutorchJsi } from '../../native/bridge';
 
 declare const tokenizerBrand: unique symbol;
@@ -8,43 +13,65 @@ declare const tokenizerBrand: unique symbol;
  * @category Types
  */
 export type Tokenizer = {
-  /** Absolute local path of the loaded `tokenizer.json`. */
+  /** Absolute local file path of the loaded `tokenizer.json`. */
   readonly path: string;
 
   /**
    * Encodes a string into token ids (special tokens are added according to the
-   * tokenizer.json post_processor).
-   * @param text The input text to tokenize.
+   * `tokenizer.json` post_processor).
+   * @param text The input text string to tokenize.
    * @returns The encoded token ids as an `Int32Array`.
+   * @throws {RnExecuTorchError} With code `EXECUTION_FAILED` if tokenization
+   * fails, `RESOURCE_BUSY` if the tokenizer is in use, or `RESOURCE_DISPOSED`
+   * if disposed.
    */
   encode(text: string): Int32Array;
 
   /**
    * Decodes token ids back into a string.
-   * @param tokens The token ids to decode, as returned by {@link encode}.
+   * @param tokens The token ids to decode (as an `Int32Array`).
    * @param skipSpecialTokens Whether to omit special tokens. Defaults to `true`.
-   * @returns The decoded text.
+   * @returns The decoded text string.
+   * @throws {RnExecuTorchError} With code `EXECUTION_FAILED` if decoding fails,
+   * `RESOURCE_BUSY` if the tokenizer is in use, or `RESOURCE_DISPOSED` if
+   * disposed.
    */
   decode(tokens: Int32Array, skipSpecialTokens?: boolean): string;
 
   /**
-   * @returns The size of the tokenizer's vocabulary.
+   * Returns the total vocabulary size.
+   * @returns The total number of tokens in the vocabulary.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the tokenizer is
+   * in use, or `RESOURCE_DISPOSED` if disposed.
    */
   getVocabSize(): number;
 
   /**
+   * Converts a numeric token id to its string piece representation.
    * @param id The token id to look up.
-   * @returns The token string for the given id.
+   * @returns The string piece for the given token id.
+   * @throws {RnExecuTorchError} With code `EXECUTION_FAILED` if id lookup
+   * fails, `RESOURCE_BUSY` if the tokenizer is in use, or `RESOURCE_DISPOSED`
+   * if disposed.
    */
   idToToken(id: number): string;
 
   /**
-   * @param token The token string to look up.
-   * @returns The id for the given token string.
+   * Converts a string piece token to its numeric token id.
+   * @param token The token piece string to look up.
+   * @returns The numeric id for the given token piece.
+   * @throws {RnExecuTorchError} With code `EXECUTION_FAILED` if token lookup
+   * fails, `RESOURCE_BUSY` if the tokenizer is in use, or `RESOURCE_DISPOSED`
+   * if disposed.
    */
   tokenToId(token: string): number;
 
-  /** Releases the native tokenizer. The instance must not be used afterwards. */
+  /**
+   * Releases the native tokenizer resources. The instance must not be used
+   * afterwards.
+   * @throws {RnExecuTorchError} With code `RESOURCE_DISPOSED` if the tokenizer
+   * has already been disposed.
+   */
   dispose(): void;
 
   /**
@@ -58,7 +85,9 @@ export type Tokenizer = {
  * Loads a HuggingFace tokenizer from a local `tokenizer.json` file.
  * @category Typescript API
  * @param tokenizerPath Absolute local path to a `tokenizer.json` file.
- * @returns The loaded tokenizer.
+ * @returns The loaded native {@link Tokenizer} instance.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the tokenizer file
+ * fails to load or parse.
  */
 export function loadTokenizer(tokenizerPath: string): Tokenizer {
   'worklet';
diff --git a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
index d0194a3c95..1065aa4cd8 100644
--- a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
@@ -1,3 +1,8 @@
+/**
+ * BIOES grammar construction, Viterbi decoding, and PII entity span extraction.
+ * @module NLP/Utils/PrivacyFilter
+ */
+
 // Privacy-filter-specific helpers: BIOES grammar construction and constrained
 // Viterbi decoding over per-token logits, plus span extraction. The decoder is
 // our own implementation of the standard Viterbi algorithm, constrained to
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
index 97cb860246..099651ebde 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
@@ -215,12 +215,6 @@ export async function createFsmnVoiceActivityDetector(
   /**
    * Synchronous version of {@link detectVoice} to be executed directly on the caller
    * or worklet thread.
-   * @param waveform The input audio samples (16 kHz mono Float32 PCM).
-   * @param options Optional per-call overrides of the detection thresholds.
-   * See {@link VadOptions}.
-   * @returns The detected speech segments, in seconds.
-   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
-   * use, or `RESOURCE_DISPOSED` if disposed.
    */
   detectVoiceWorklet: (waveform: Float32Array, options?: VadOptions) => Segment[];
 
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
index 837190bc78..958112a7e0 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
@@ -145,14 +145,6 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
   /**
    * Synchronous version of {@link transcribe} to be executed directly on the
    * caller or worklet thread.
-   * @param audio Raw 16 kHz mono PCM audio samples (Float32Array).
-   * @param options Transcription options. See {@link WhisperSttOptions}.
-   * @param onToken Optional callback fired on the RN thread for each decoded
-   * token.
-   * @returns The full transcript string.
-   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the language is
-   * unsupported, `RESOURCE_BUSY` if the model is in use, or `RESOURCE_DISPOSED`
-   * if disposed.
    */
   transcribeWorklet: (
     audio: Float32Array,

From 051b30ea851d5aac339b904d194b5823b07cca63 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 18:04:06 +0200
Subject: [PATCH 15/38] docs(llm): standardize TSDoc annotations, module
 headers, and error throws

---
 .../src/extensions/llm/index.ts                       |  6 ++++++
 .../src/extensions/llm/llmRunner.ts                   | 11 ++++++++---
 .../src/extensions/llm/tasks/llmChatSession.ts        |  6 ++++++
 .../src/extensions/llm/utils/chatPreprocessor.ts      |  6 ++++++
 .../src/extensions/llm/utils/tokenizerConfig.ts       |  6 ++++++
 .../src/extensions/llm/utils/toolCalling.ts           |  6 ++++++
 6 files changed, 38 insertions(+), 3 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/llm/index.ts b/packages/react-native-executorch/src/extensions/llm/index.ts
index ba276aba05..fe61f5f582 100644
--- a/packages/react-native-executorch/src/extensions/llm/index.ts
+++ b/packages/react-native-executorch/src/extensions/llm/index.ts
@@ -1,3 +1,9 @@
+/**
+ * Large Language Model (LLM) runners, multi-turn chat sessions, and
+ * tool-calling utilities.
+ * @module LLM
+ */
+
 export * from './llmRunner';
 export * from './utils/chatPreprocessor';
 export * from './utils/tokenizerConfig';
diff --git a/packages/react-native-executorch/src/extensions/llm/llmRunner.ts b/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
index 6cca8b176f..4a93503af4 100644
--- a/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
+++ b/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
@@ -1,3 +1,8 @@
+/**
+ * Low-level native ExecuTorch LLM runner types and factory.
+ * @module LLM/LLMRunner
+ */
+
 import type { Tensor } from '../../core/tensor';
 import { rnexecutorchJsi } from '../../native/bridge';
 
@@ -153,9 +158,9 @@ export type LLMRunner = {
  * @category Typescript API
  * @param modelPath Path to the local `.pte` model file.
  * @param tokenizerPath Path to the local tokenizer configuration file (e.g. `tokenizer.json`).
- * @param modalities List of supported input non-text modalities (e.g. `['image']`).
- * Defaults to text-only.
- * @returns A native LLMRunner instance.
+ * @param modalities List of supported input non-text modalities (e.g.
+ * `['image']`). When omitted, defaults to text-only.
+ * @returns A native {@link LLMRunner} instance.
  */
 export function createLLMRunner(
   modelPath: string,
diff --git a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
index f936514aef..6efa38333f 100644
--- a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
+++ b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
@@ -1,3 +1,9 @@
+/**
+ * Multi-turn LLM chat session with history management, tool calling, and KV
+ * cache prefilling.
+ * @module LLM/Tasks/LLMChatSession
+ */
+
 import { scheduleOnRN, type WorkletRuntime } from 'react-native-worklets';
 import RNBlobUtil from 'react-native-blob-util';
 
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
index 2e33654522..7ebe60d529 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
@@ -1,3 +1,9 @@
+/**
+ * Multimodal chat message template rendering, incremental prompt diffing, and
+ * media tensor preprocessing for LLM runners.
+ * @module LLM/Utils/ChatPreprocessor
+ */
+
 import { Template } from '@huggingface/jinja';
 
 import { tensor, type Tensor } from '../../../core/tensor';
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
index ce13438868..c11b2454f5 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
@@ -1,3 +1,9 @@
+/**
+ * Parser for HuggingFace `tokenizer_config.json` chat templates and special
+ * tokens.
+ * @module LLM/Utils/TokenizerConfig
+ */
+
 import { RnExecuTorchError } from '../../../core/error';
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
index d591fa4167..e2b413f297 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
@@ -1,3 +1,9 @@
+/**
+ * Tool call definition, XML/JSON parsing, and schema generation for LLM
+ * function calling.
+ * @module LLM/Utils/ToolCalling
+ */
+
 import type { ChatMessageContent } from './chatPreprocessor';
 
 /**

From 801ecb23c1d081cbac25a5886ca357fcba1b2aac Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 18:39:40 +0200
Subject: [PATCH 16/38] fix: improve JSDoc/TSDoc across src/ (25 files)

- Add missing @category tags (schema.ts, telemetry.ts)
- Add field-level JSDoc on LLM types (toolCalling, chatPreprocessor, tokenizerConfig)
- Add @throws tags to LLM functions that throw (chatPreprocessor, tokenizerConfig, llmChatSession)
- Add missing docs on schema.ts convenience functions (f32/i64/i32/ui8/bool, constr.eq/linear)
- Standardize worklet one-liner pattern (semanticSegmentation, privacyFilter)
- Improve hook @returns to reference create<Task> via @link (all 15 hooks)
- Misc: add module docs (ops/index, fetcher/index), fix @link in vadUtils, remove @category from private postprocess
---
 .../src/core/schema.ts                        | 51 ++++++++++++++++++-
 .../src/extensions/cv/ops/index.ts            |  4 ++
 .../extensions/cv/tasks/keypointDetection.ts  |  1 -
 .../cv/tasks/semanticSegmentation.ts          |  4 +-
 .../extensions/llm/tasks/llmChatSession.ts    |  4 +-
 .../extensions/llm/utils/chatPreprocessor.ts  |  8 +++
 .../extensions/llm/utils/tokenizerConfig.ts   |  4 ++
 .../src/extensions/llm/utils/toolCalling.ts   | 13 +++++
 .../src/extensions/nlp/tasks/privacyFilter.ts |  3 +-
 .../src/fetcher/telemetry.ts                  |  2 +
 .../src/hooks/useClassifier.ts                |  4 +-
 .../src/hooks/useImageEmbedder.ts             |  4 +-
 .../src/hooks/useInstanceSegmenter.ts         |  4 +-
 .../src/hooks/useKeypointDetector.ts          |  4 +-
 .../src/hooks/useLLMChatSession.ts            |  6 +--
 .../src/hooks/useObjectDetector.ts            |  4 +-
 .../src/hooks/usePrivacyFilter.ts             |  4 +-
 .../src/hooks/useSemanticSegmenter.ts         |  4 +-
 .../src/hooks/useSpeechToText.ts              |  4 +-
 .../src/hooks/useStyleTransfer.ts             |  4 +-
 .../src/hooks/useTextEmbedder.ts              |  4 +-
 .../src/hooks/useTextToImage.ts               |  4 +-
 .../src/hooks/useTextToSpeech.ts              |  8 +--
 .../src/hooks/useTokenizer.ts                 |  4 +-
 .../src/hooks/useVoiceActivityDetector.ts     |  4 +-
 25 files changed, 121 insertions(+), 39 deletions(-)

diff --git a/packages/react-native-executorch/src/core/schema.ts b/packages/react-native-executorch/src/core/schema.ts
index 8cd52d9bc5..a040788991 100644
--- a/packages/react-native-executorch/src/core/schema.ts
+++ b/packages/react-native-executorch/src/core/schema.ts
@@ -337,17 +337,64 @@ export const SymbolicTensor = (dtype: DType, shape: SymbolicShape) => {
   return { kind: 'Tensor', dtype, shape: typedShape } as TensorSpec<SymbolicDim>;
 };
 
+/**
+ * Shorthand for `SymbolicTensor('float32', shape)`.
+ * @category Typescript API
+ * @param shape Dimension sizes of the tensor.
+ * @returns A {@link SymbolicTensor} with `float32` data type.
+ */
 export const f32 = (...shape: SymbolicShape) => SymbolicTensor('float32', shape);
+/**
+ * Shorthand for `SymbolicTensor('int64', shape)`.
+ * @category Typescript API
+ * @param shape Dimension sizes of the tensor.
+ * @returns A {@link SymbolicTensor} with `int64` data type.
+ */
 export const i64 = (...shape: SymbolicShape) => SymbolicTensor('int64', shape);
+/**
+ * Shorthand for `SymbolicTensor('int32', shape)`.
+ * @category Typescript API
+ * @param shape Dimension sizes of the tensor.
+ * @returns A {@link SymbolicTensor} with `int32` data type.
+ */
 export const i32 = (...shape: SymbolicShape) => SymbolicTensor('int32', shape);
+/**
+ * Shorthand for `SymbolicTensor('uint8', shape)`.
+ * @category Typescript API
+ * @param shape Dimension sizes of the tensor.
+ * @returns A {@link SymbolicTensor} with `uint8` data type.
+ */
 export const ui8 = (...shape: SymbolicShape) => SymbolicTensor('uint8', shape);
+/**
+ * Shorthand for `SymbolicTensor('bool', shape)`.
+ * @category Typescript API
+ * @param shape Dimension sizes of the tensor.
+ * @returns A {@link SymbolicTensor} with `bool` data type.
+ */
 export const bool = (...shape: SymbolicShape) => SymbolicTensor('bool', shape);
 
-/** Helper namespace for declaring runtime constraints. */
+/**
+ * Helper namespace for declaring runtime constraints.
+ * @category Typescript API
+ */
 export const constr = {
+  /**
+   * Declares that all given dimensions must be equal at runtime.
+   * @param dims Dimensions that must share the same concrete value.
+   * @returns An {@link EqualityConstraint} across the given dimensions.
+   */
   eq: (...dims: DimRef[]): EqualityConstraint => {
     return { kind: 'equality', dims };
   },
+  /**
+   * Declares a linear relation between two dimensions: `dimLhs = a * dimRhs +
+   * b`.
+   * @param dimLhs The left-hand-side dimension.
+   * @param dimRhs The right-hand-side dimension.
+   * @param a Slope coefficient.
+   * @param b Intercept coefficient (defaults to `0`).
+   * @returns A {@link LinearConstraint} relating the two dimensions.
+   */
   linear: (dimLhs: DimRef, dimRhs: DimRef, a: number, b: number = 0): LinearConstraint => {
     return { kind: 'linear', dimLhs, dimRhs, coefficients: [a, b] };
   },
@@ -356,6 +403,7 @@ export const constr = {
 /**
  * Constructs a method specification mapping a method name to its parameter
  * specs and runtime constraints.
+ * @category Typescript API
  * @param name The execution method name (e.g., `'forward'`).
  * @param inputs Ordered list of parameter specifications for method inputs.
  * @param outputs Ordered list of parameter specifications for method outputs.
@@ -891,6 +939,7 @@ function createSpecMatch<K extends PropertyKey>(
  * accessors.
  * @throws {RnExecuTorchError} With code `SCHEMA_MISMATCH`, describing
  * why every variant failed.
+ * @category Typescript API
  */
 export function validateSpec<const T extends Record<string, ModelSpec<SymbolicDim>>>(
   exportedModelSpec: ModelSpec<ConcreteDim>,
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/index.ts b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
index 4128d25527..ce242853b9 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
@@ -1,3 +1,7 @@
+/**
+ * Aggregated image, box, and point operations.
+ * @module CV Ops
+ */
 export * as image from './image';
 export * as boxes from './boxes';
 export * as points from './points';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index 004fe4f366..b7658b695c 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -96,7 +96,6 @@ export type KeypointDetection<F extends BoxFormat, L extends PropertyKey> = {
 /**
  * Post-processes model outputs by applying Non-Maximum Suppression (NMS) and
  * scaling coordinates.
- * @category Utils
  * @param tBoxes Bounding boxes tensor output from inference.
  * @param tScores Scores tensor output from inference.
  * @param tKeypoints Keypoints tensor output from inference.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
index 1731817c98..1c88186016 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
@@ -136,8 +136,8 @@ export async function createSemanticSegmenter<L extends PropertyKey = string>(
   ) => Promise<SemanticSegmentationResult<L>>;
 
   /**
-   * Runs semantic segmentation synchronously.
-   * @see {@link segment} for details.
+   * Synchronous version of {@link segment} to be executed directly on the
+   * caller or worklet thread.
    */
   segmentWorklet: (
     input: ImageBuffer,
diff --git a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
index 6efa38333f..9f8acf4e14 100644
--- a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
+++ b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
@@ -144,6 +144,8 @@ export type LLMChatSession = {
    * @param onToken Callback fired on the RN thread for each decoded token.
    * @param genConfig Generation options overriding session defaults.
    * @returns A promise resolving to the generated messages and turn stats.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the message is
+   * malformed, or `INVALID_STATE` if the chat template is non-monotonic.
    */
   sendMessage(
     message: ChatMessageContent,
@@ -188,7 +190,7 @@ const DEFAULT_MAX_TURNS = 5;
  * @param config Model configuration containing model, tokenizer, and tokenizer config paths.
  * @param options Custom generation, tool calling, and state options.
  * @param runtime The worklet runtime thread to run native generation on.
- * @returns A Promise resolving to an LLMChatSession instance.
+ * @returns A Promise resolving to an {@link LLMChatSession} instance.
  */
 export async function createLLMChatSession(
   config: LLMModel,
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
index 7ebe60d529..531d92f1ab 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
@@ -68,7 +68,9 @@ export type LLMAudioPreprocessorConfig = {
  * @category Types
  */
 export type LLMMediaPreprocessorConfig = {
+  /** Image preprocessing configuration for vision-language models. */
   readonly image?: LLMImagePreprocessorConfig;
+  /** Audio preprocessing configuration for audio-language models. */
   readonly audio?: LLMAudioPreprocessorConfig;
 };
 
@@ -77,9 +79,13 @@ export type LLMMediaPreprocessorConfig = {
  * @category Types
  */
 export type ChatPreprocessorConfig = {
+  /** Jinja chat template string for prompt rendering. */
   readonly chatTemplate: string;
+  /** Tool definitions available to the model. */
   readonly tools?: readonly ToolDefinition[];
+  /** Supported input modalities (e.g. `['image']`). */
   readonly modalities?: readonly Modality[];
+  /** Media preprocessing configuration for non-text inputs. */
   readonly preprocessorConfig?: LLMMediaPreprocessorConfig;
 };
 
@@ -167,6 +173,8 @@ const MEDIA_SENTINEL_REGEX = /\uFFFC__ET_MEDIA_(\d+)__/g;
  * @param content Chat message text or media input array.
  * @param options Sentinel tokens, supported modalities, and starting media index.
  * @returns Object containing serialized text and a map of media items.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if a media kind is
+ * unsupported or a required sentinel token is missing.
  */
 function chatContentToString(
   content?: ChatMessageContent,
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
index c11b2454f5..6a5cba0c86 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
@@ -11,7 +11,9 @@ import { RnExecuTorchError } from '../../../core/error';
  * @category Types
  */
 export type TokenizerChatConfig = {
+  /** Jinja chat template string for prompt rendering. */
   readonly chatTemplate: string;
+  /** End-of-sequence token string. */
   readonly eosToken: string;
 };
 
@@ -28,6 +30,8 @@ function resolveToken(token: unknown): string | undefined {
  * @category Utils
  * @param config Raw JSON object from tokenizer_config.json.
  * @returns A parsed TokenizerChatConfig object.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if `chat_template` is not
+ * a string or `eos_token` is missing.
  */
 export function parseTokenizerConfig(config: any): TokenizerChatConfig {
   let chatTemplate = config.chat_template;
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
index e2b413f297..0ab0d4c00b 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
@@ -11,8 +11,11 @@ import type { ChatMessageContent } from './chatPreprocessor';
  * @category Types
  */
 export type ToolParameters = {
+  /** JSON Schema type, typically `'object'`. */
   readonly type?: 'object' | string;
+  /** JSON Schema properties mapping parameter names to their schemas. */
   readonly properties?: Record<string, unknown>;
+  /** Names of required parameters. */
   readonly required?: readonly string[];
   readonly [key: string]: unknown;
 };
@@ -22,10 +25,14 @@ export type ToolParameters = {
  * @category Types
  */
 export type ToolDefinition<Args extends Record<string, unknown> = Record<string, unknown>> = {
+  /** Tool type discriminator, typically `'function'`. */
   readonly type: 'function' | string;
   readonly function: {
+    /** Function name the model should invoke. */
     readonly name: string;
+    /** Human-readable description of what the tool does. */
     readonly description?: string;
+    /** JSON Schema describing the function's parameters. */
     readonly parameters?: ToolParameters;
     readonly [key: string]: unknown;
   };
@@ -41,10 +48,14 @@ export type ToolDefinition<Args extends Record<string, unknown> = Record<string,
  * @category Types
  */
 export type ToolCall = {
+  /** Optional tool call identifier for matching responses. */
   readonly id?: string;
+  /** Tool type discriminator, typically `'function'`. */
   readonly type?: 'function' | string;
   readonly function: {
+    /** Name of the function being called. */
     readonly name: string;
+    /** Parsed arguments passed to the function. */
     readonly arguments: Record<string, unknown>;
   };
 };
@@ -54,7 +65,9 @@ export type ToolCall = {
  * @category Types
  */
 export type ToolParserResult = {
+  /** Detected tool calls extracted from the model output. */
   readonly toolCalls: readonly ToolCall[];
+  /** Remaining text content after tool call extraction. */
   readonly textContent?: string;
 };
 
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
index aeaecbcccd..72598ff805 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
@@ -117,7 +117,8 @@ export async function createPrivacyFilter<Label extends string>(
   detectPii: (input: string) => Promise<PiiEntity<PiiEntityType<Label>>[]>;
 
   /**
-   * Synchronous version of {@link detectPii}.
+   * Synchronous version of {@link detectPii} to be executed directly on the
+   * caller or worklet thread.
    */
   detectPiiWorklet: (input: string) => PiiEntity<PiiEntityType<Label>>[];
 }> {
diff --git a/packages/react-native-executorch/src/fetcher/telemetry.ts b/packages/react-native-executorch/src/fetcher/telemetry.ts
index 6dfda5ce0f..0aa9fdb26a 100644
--- a/packages/react-native-executorch/src/fetcher/telemetry.ts
+++ b/packages/react-native-executorch/src/fetcher/telemetry.ts
@@ -42,6 +42,7 @@ function isSwmHuggingFaceRepo(url: URL): boolean {
  * download-stats convention. No-op for any other host.
  *
  * See https://huggingface.co/docs/hub/models-download-stats
+ * @category Utils
  * @param uri The URI of the file being downloaded.
  */
 export function triggerHuggingFaceDownloadCounter(uri: string): void {
@@ -97,6 +98,7 @@ function getModelNameFromUri(uri: string): string {
  * Sends an anonymous download event to the Software Mansion analytics endpoint,
  * unless the app has opted out via {@link setTelemetryEnabled}. Fire-and-forget;
  * never throws and never blocks the download.
+ * @category Utils
  * @param uri The URI of the downloaded resource.
  */
 export function triggerDownloadEvent(uri: string): void {
diff --git a/packages/react-native-executorch/src/hooks/useClassifier.ts b/packages/react-native-executorch/src/hooks/useClassifier.ts
index c129d7d0bc..a2e2efec03 100644
--- a/packages/react-native-executorch/src/hooks/useClassifier.ts
+++ b/packages/react-native-executorch/src/hooks/useClassifier.ts
@@ -16,8 +16,8 @@ import { createClassifier, type ClassifierModel } from '../extensions/cv/tasks/c
  * @param config The image classification model configuration.
  * See {@link ClassifierModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, labels, and classification functions.
+ * @returns The same object as {@link createClassifier} (without `dispose`),
+ * combined with loading state, download progress, and labels.
  * @see {@link createClassifier}
  */
 export function useClassifier<L>(config: ClassifierModel<L>, options?: ResourceOptions) {
diff --git a/packages/react-native-executorch/src/hooks/useImageEmbedder.ts b/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
index f9669ee0a4..01cabea0ec 100644
--- a/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
+++ b/packages/react-native-executorch/src/hooks/useImageEmbedder.ts
@@ -18,8 +18,8 @@ import {
  * @param config The image embedder model configuration.
  * See {@link ImageEmbedderModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and embedding functions.
+ * @returns The same object as {@link createImageEmbedder} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createImageEmbedder}
  */
 export function useImageEmbedder(config: ImageEmbedderModel, options?: ResourceOptions) {
diff --git a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
index a9bd55bb7d..e91c015767 100644
--- a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
@@ -21,8 +21,8 @@ import {
  * @param config The instance segmentation model configuration.
  * See {@link InstanceSegmenterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, labels, and instance segmentation functions.
+ * @returns The same object as {@link createInstanceSegmenter} (without `dispose`),
+ * combined with loading state, download progress, and labels.
  * @see {@link createInstanceSegmenter}
  */
 export function useInstanceSegmenter<F extends BoxFormat, L>(
diff --git a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
index 58ba581d97..89ec0012f8 100644
--- a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
@@ -21,8 +21,8 @@ import {
  * @param config The keypoint detection model configuration.
  * See {@link KeypointDetectorModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, landmarks, and keypoint detection functions.
+ * @returns The same object as {@link createKeypointDetector} (without `dispose`),
+ * combined with loading state, download progress, and landmarks.
  * @see {@link createKeypointDetector}
  */
 export function useKeypointDetector<F extends BoxFormat, L extends PropertyKey>(
diff --git a/packages/react-native-executorch/src/hooks/useLLMChatSession.ts b/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
index 6b93393596..46c358fd2c 100644
--- a/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
+++ b/packages/react-native-executorch/src/hooks/useLLMChatSession.ts
@@ -18,9 +18,9 @@ import {
  * @category Hooks
  * @param config The LLM model configuration. See {@link LLMModel}.
  * @param options Chat session options and load/caching options.
- * See {@link ResourceOptions}.
- * @returns An object containing the session's loading state, error, download
- * progress, and chat functions.
+ * See {@link LLMChatSessionOptions} & {@link ResourceOptions}.
+ * @returns The same object as {@link createLLMChatSession} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createLLMChatSession}
  */
 export function useLLMChatSession(
diff --git a/packages/react-native-executorch/src/hooks/useObjectDetector.ts b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
index 17241767b1..4acbccd4da 100644
--- a/packages/react-native-executorch/src/hooks/useObjectDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
@@ -21,8 +21,8 @@ import {
  * @param config The object detection model configuration.
  * See {@link ObjectDetectorModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, labels, and object detection functions.
+ * @returns The same object as {@link createObjectDetector} (without `dispose`),
+ * combined with loading state, download progress, and labels.
  * @see {@link createObjectDetector}
  */
 export function useObjectDetector<F extends BoxFormat, L>(
diff --git a/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts b/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
index 47a567408b..ef2fc25784 100644
--- a/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
+++ b/packages/react-native-executorch/src/hooks/usePrivacyFilter.ts
@@ -20,8 +20,8 @@ import {
  * @param config The privacy filter model configuration.
  * See {@link PrivacyFilterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and detection functions.
+ * @returns The same object as {@link createPrivacyFilter} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createPrivacyFilter}
  */
 export function usePrivacyFilter<Label extends string>(
diff --git a/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts b/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
index c37800ca30..a71ff66321 100644
--- a/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useSemanticSegmenter.ts
@@ -19,8 +19,8 @@ import {
  * @param config The semantic segmentation model configuration.
  * See {@link SemanticSegmenterModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, labels, and semantic segmentation functions.
+ * @returns The same object as {@link createSemanticSegmenter} (without `dispose`),
+ * combined with loading state, download progress, and labels.
  * @see {@link createSemanticSegmenter}
  */
 export function useSemanticSegmenter<L extends PropertyKey = string>(
diff --git a/packages/react-native-executorch/src/hooks/useSpeechToText.ts b/packages/react-native-executorch/src/hooks/useSpeechToText.ts
index 75380a9d3c..9dc60e092a 100644
--- a/packages/react-native-executorch/src/hooks/useSpeechToText.ts
+++ b/packages/react-native-executorch/src/hooks/useSpeechToText.ts
@@ -19,8 +19,8 @@ import {
  * @param config The Whisper speech-to-text model configuration.
  * See {@link WhisperSttModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and transcription functions.
+ * @returns The same object as {@link createWhisperSpeechToText} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createWhisperSpeechToText}
  */
 export function useSpeechToText<L extends WhisperLanguage = WhisperLanguage>(
diff --git a/packages/react-native-executorch/src/hooks/useStyleTransfer.ts b/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
index e9780cb6bd..a986494a14 100644
--- a/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
+++ b/packages/react-native-executorch/src/hooks/useStyleTransfer.ts
@@ -15,8 +15,8 @@ import { createStyleTransfer, type StyleTransferModel } from '../extensions/cv/t
  * @param config The style transfer model configuration.
  * See {@link StyleTransferModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and style transfer functions.
+ * @returns The same object as {@link createStyleTransfer} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createStyleTransfer}
  */
 export function useStyleTransfer(config: StyleTransferModel, options?: ResourceOptions) {
diff --git a/packages/react-native-executorch/src/hooks/useTextEmbedder.ts b/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
index 779980ec9e..ee0f64e4d5 100644
--- a/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
+++ b/packages/react-native-executorch/src/hooks/useTextEmbedder.ts
@@ -15,8 +15,8 @@ import { createTextEmbedder, type TextEmbedderModel } from '../extensions/nlp/ta
  * @param config The text embedder model configuration.
  * See {@link TextEmbedderModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and embedding functions.
+ * @returns The same object as {@link createTextEmbedder} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createTextEmbedder}
  */
 export function useTextEmbedder(config: TextEmbedderModel, options?: ResourceOptions) {
diff --git a/packages/react-native-executorch/src/hooks/useTextToImage.ts b/packages/react-native-executorch/src/hooks/useTextToImage.ts
index a4a5646ad5..947274a6ca 100644
--- a/packages/react-native-executorch/src/hooks/useTextToImage.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToImage.ts
@@ -17,8 +17,8 @@ import {
  * @category Hooks
  * @param config The SDXS model configuration. See {@link SdxsTextToImageModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and image generation functions.
+ * @returns The same object as {@link createSdxsTextToImage} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createSdxsTextToImage}
  */
 export function useTextToImage(config: SdxsTextToImageModel, options?: ResourceOptions) {
diff --git a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
index f94f6ea030..9c8dc16871 100644
--- a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
@@ -52,8 +52,8 @@ export type TtsHookResult<
  * @typeParam K Voice keys record constraint.
  * @param config The Kokoro TTS model configuration. See {@link KokoroTtsModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and speech synthesis functions.
+ * @returns The same object as {@link createKokoroTextToSpeech} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createKokoroTextToSpeech}
  */
 export function useTextToSpeech<K extends PropertyKey>(
@@ -75,8 +75,8 @@ export function useTextToSpeech<K extends PropertyKey>(
  * @param config The Supertonic TTS model configuration.
  * See {@link SupertonicTtsModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, and speech synthesis functions.
+ * @returns The same object as {@link createSupertonicTextToSpeech} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createSupertonicTextToSpeech}
  */
 export function useTextToSpeech<K extends PropertyKey>(
diff --git a/packages/react-native-executorch/src/hooks/useTokenizer.ts b/packages/react-native-executorch/src/hooks/useTokenizer.ts
index 4c949e06d1..89621b8506 100644
--- a/packages/react-native-executorch/src/hooks/useTokenizer.ts
+++ b/packages/react-native-executorch/src/hooks/useTokenizer.ts
@@ -14,8 +14,8 @@ import { createTokenizer } from '../extensions/nlp/tasks/tokenization';
  * @category Hooks
  * @param tokenizerPath A remote URL or local path to a `tokenizer.json` file.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the tokenizer's loading state, error, download
- * progress, and tokenization functions.
+ * @returns The same object as {@link createTokenizer} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createTokenizer}
  */
 export function useTokenizer(tokenizerPath: string, options?: ResourceOptions) {
diff --git a/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts b/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts
index d0ef117b59..2765a43264 100644
--- a/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useVoiceActivityDetector.ts
@@ -17,8 +17,8 @@ import {
  * @category Hooks
  * @param config The VAD model configuration. See {@link FsmnVadModel}.
  * @param options Load and caching options. See {@link ResourceOptions}.
- * @returns An object containing the model's loading state, error, download
- * progress, one-shot detection functions, and live detection controls.
+ * @returns The same object as {@link createFsmnVoiceActivityDetector} (without `dispose`),
+ * combined with loading state and download progress.
  * @see {@link createFsmnVoiceActivityDetector}
  */
 export function useVoiceActivityDetector(config: FsmnVadModel, options?: ResourceOptions) {

From c780d656c22e1d24533943de09ba35189cca0d6e Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Fri, 21 Aug 2026 18:43:11 +0200
Subject: [PATCH 17/38] fix: add missing @link for RnExecuTorchError in
 wrapAsync @returns

---
 packages/react-native-executorch/src/core/runtime.ts | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/packages/react-native-executorch/src/core/runtime.ts b/packages/react-native-executorch/src/core/runtime.ts
index 6c1f349061..6525dee6d6 100644
--- a/packages/react-native-executorch/src/core/runtime.ts
+++ b/packages/react-native-executorch/src/core/runtime.ts
@@ -43,7 +43,7 @@ export const defaultWorkletRuntime = createWorkletRuntime({
  * @param runtime The worklet runtime to dispatch `fn` to. Defaults to
  * {@link defaultWorkletRuntime}.
  * @returns An async function with the same signature as `fn` that resolves to
- * `fn`'s return value or rejects with an `RnExecuTorchError` if `fn` throws.
+ * `fn`'s return value or rejects with an {@link RnExecuTorchError} if `fn` throws.
  * @throws {RnExecuTorchError} Propagates any error thrown inside `fn` across the
  * worklet thread boundary.
  * @example

From e3c4c8b3d3d3850679a841a0d9457e2839f75f0a Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 13:06:33 +0200
Subject: [PATCH 18/38] docs(ocr): standardize TSDoc annotations, extract task
 options, and document quad geometry

---
 .../src/extensions/cv/ops/quad.ts             | 40 +++++++++++++------
 .../src/extensions/cv/tasks/paddleOcr.ts      | 39 ++++++++++++------
 .../src/extensions/cv/utils/paddleOcrUtils.ts | 13 ++++--
 packages/react-native-executorch/src/index.ts | 16 ++++----
 .../react-native-executorch/src/models.ts     |  3 ++
 5 files changed, 75 insertions(+), 36 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
index 1955e11cb7..09e6299a1f 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -1,3 +1,9 @@
+/**
+ * Quadrilateral geometry, bounding box extraction, and image rectification
+ * utilities.
+ * @module CV/Ops/Quad
+ */
+
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import { RnExecuTorchError } from '../../../core/error';
@@ -22,6 +28,8 @@ export type Quad = readonly [Point, Point, Point, Point];
  * @param points The points to enclose.
  * @param format The coordinate format of the returned box.
  * @returns The enclosing {@link BoundingBox} in `format`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the bounding box
+ * format is unsupported.
  */
 export function boundingBoxOfPoints<F extends BoxFormat>(
   points: readonly Point[],
@@ -103,24 +111,29 @@ export function quadSize(ordered: Quad): { width: number; height: number } {
   return { width, height };
 }
 
+/**
+ * Configuration options for scaling quad coordinates.
+ * @category Types
+ */
+export type ScaleQuadOptions = {
+  /** The source bounds (e.g. model input dimensions). */
+  readonly from: { readonly width: number; readonly height: number };
+  /** The destination bounds (e.g. original image dimensions). */
+  readonly to: { readonly width: number; readonly height: number };
+  /** The mode used to resize the image (excluding `'crop'`). */
+  readonly resizeMode?: Exclude<ResizeMode, 'crop'>;
+};
+
 /**
  * Rescales a quad from one frame to another, clamping the result to the target
  * bounds. The counterpart of {@link scaleBox} for quads.
  * @category Typescript API
  * @param quad The quad, expressed in the `from` frame.
- * @param options `from` is the frame the quad is expressed in, `to` the frame to
- * express it in, and `resizeMode` how the two were fitted (default
- * `'letterbox'`).
+ * @param options Options detailing the scaling factors and resize mode.
+ * See {@link ScaleQuadOptions}.
  * @returns The four corners in `to` pixels.
  */
-export function scaleQuad(
-  quad: Quad,
-  options: {
-    readonly from: { readonly width: number; readonly height: number };
-    readonly to: { readonly width: number; readonly height: number };
-    readonly resizeMode?: Exclude<ResizeMode, 'crop'>;
-  }
-): Quad {
+export function scaleQuad(quad: Quad, options: ScaleQuadOptions): Quad {
   'worklet';
   const { from, to, resizeMode } = options;
   const map = (p: Point): Point => {
@@ -153,8 +166,11 @@ export type RectifyQuadOptions = {
  * @param dst The pre-allocated destination canvas, `uint8` `[H', W', C]`, with
  * the same channel count as `src`. Must not alias `src`.
  * @param quad The region corners (TL, TR, BR, BL) in `src` pixels.
- * @param options Content width, alignment, and padding.
+ * @param options Content width, alignment, and padding. See {@link RectifyQuadOptions}.
  * @returns The destination tensor `dst`.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shapes,
+ * data types, or quadrilateral coordinates are invalid, `RESOURCE_BUSY` if a
+ * tensor is in use, or `RESOURCE_DISPOSED` if either tensor was disposed.
  */
 export function rectifyQuad(
   src: Tensor,
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 1b7482499e..366b882bd2 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -1,7 +1,8 @@
-// PP-OCRv6: a DBNet text detector and an SVTR recognizer fused into one PTE.
-// One pass locates every text line on the page, warps each to the recognizer
-// canvas and reads it. Worklet source order matters here: a referenced worklet
-// must be defined above its callers.
+/**
+ * PP-OCRv6 optical character recognition (DBNet text detection + SVTR text
+ * recognition).
+ * @module CV/Tasks/PaddleOcr
+ */
 
 import type { WorkletRuntime } from 'react-native-worklets';
 import RNBlobUtil from 'react-native-blob-util';
@@ -70,6 +71,18 @@ export type PaddleOcrModelOptions = {
   readonly defaultConfidenceThreshold: number;
 };
 
+/**
+ * Optional configuration parameters for optical character recognition inference.
+ * @category Types
+ */
+export type RecognizeCharactersOptions = {
+  /**
+   * Minimum confidence threshold to retain recognized text regions, in `[0, 1]`.
+   * Overrides the model's `defaultConfidenceThreshold` for this call.
+   */
+  readonly confidenceThreshold?: number;
+};
+
 /**
  * Model configuration for the PP-OCRv6 pipeline: one fused detect/recognize PTE,
  * the charset published beside it, and the run options.
@@ -342,13 +355,14 @@ function greedyCtcDecode(
  * warps each to the recognizer canvas and reads it, returning the lines in
  * reading order.
  * @category Typescript API
- * @param config Model path, charset path, and run options.
+ * @param config Model path, charset path, and run options. See {@link PaddleOcrModel}.
  * @param runtime Optional worklet runtime thread.
  * @returns A promise resolving to an object containing recognition and disposal
  * controls.
- * @throws {RnExecuTorchError} With code `SCHEMA_MISMATCH` if the loaded model
- * does not match the PP-OCRv6 detect/recognize contract, or if the charset does
- * not match the recognizer's vocabulary.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model or charset
+ * fails to load, `SCHEMA_MISMATCH` if the loaded model does not match the
+ * PP-OCRv6 detect/recognize contract, or if the charset does not match the
+ * recognizer's vocabulary.
  */
 export async function createPaddleOcr(
   config: PaddleOcrModel,
@@ -362,14 +376,13 @@ export async function createPaddleOcr(
   /**
    * Detects and recognizes every text line in the given image.
    * @param input The input image buffer.
-   * @param options Per-call overrides. `confidenceThreshold` replaces the
-   * model's `defaultConfidenceThreshold` for this call.
+   * @param options Per-call overrides. See {@link RecognizeCharactersOptions}.
    * @returns A promise resolving to the recognized lines in reading order
    * (leftmost column top to bottom, then the next column).
    */
   recognizeCharacters: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number }
+    options?: RecognizeCharactersOptions
   ) => Promise<OcrDetection[]>;
 
   /**
@@ -378,7 +391,7 @@ export async function createPaddleOcr(
    */
   recognizeCharactersWorklet: (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number }
+    options?: RecognizeCharactersOptions
   ) => OcrDetection[];
 }> {
   const { modelPath, charsetPath, modelOpts } = config;
@@ -430,7 +443,7 @@ export async function createPaddleOcr(
 
   const recognizeCharactersWorklet = (
     input: ImageBuffer,
-    options?: { confidenceThreshold?: number }
+    options?: RecognizeCharactersOptions
   ): OcrDetection[] => {
     'worklet';
     const confidenceThreshold =
diff --git a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
index b7f2ec3b10..4c6aa97434 100644
--- a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
@@ -1,6 +1,7 @@
-// Native decode helper for the PP-OCRv6 pipeline. Wraps a fused C++ op: tracing
-// contours in TypeScript would mean pulling the whole detector output across the
-// bridge per call.
+/**
+ * DBNet contour tracing and text quad extraction utilities for PP-OCRv6.
+ * @module CV/Utils/PaddleOcrUtils
+ */
 
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
@@ -51,8 +52,12 @@ export type DbnetDecodeOptions = {
  * survivors back to their unshrunk size.
  * @category Typescript API
  * @param probabilityMap The `detect` output, shape `[1, 1, H, W]`, post-sigmoid.
- * @param options Decode thresholds.
+ * @param options Decode thresholds. See {@link DbnetDecodeOptions}.
  * @returns The decoded quads, in detector-input pixel space and arbitrary order.
+ * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if tensor shape or
+ * data type is invalid, `RESOURCE_BUSY` if the tensor is in use,
+ * `RESOURCE_DISPOSED` if the tensor was disposed, or `EXECUTION_FAILED` if
+ * native decode returns invalid output.
  */
 export function extractDbnetTextQuads(probabilityMap: Tensor, options: DbnetDecodeOptions): Quad[] {
   'worklet';
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 9fc98cd7b1..6d57dc7fab 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -6,20 +6,22 @@
  *
  * - **React Hooks (`use<Task>`)**:
  *   Declarative hooks (e.g. {@link useClassifier}, {@link useLLMChatSession},
- *   {@link useSpeechToText}) designed for UI components. They manage
- *   downloading remote model assets with progress tracking, compilation, and
- *   automatic native memory disposal on unmount.
+ *   {@link useSpeechToText}, {@link useOpticalCharacterRecognizer}) designed
+ *   for UI components. They manage downloading remote model assets with
+ *   progress tracking, compilation, and automatic native memory disposal on
+ *   unmount.
  *
  * - **Task APIs (`create<Task>`)**:
  *   Imperative, promise-based pipelines (e.g. {@link createClassifier},
- *   {@link createLLMChatSession}, {@link createWhisperSpeechToText}) with
- *   built-in preprocessing, postprocessing, and tokenizers. Ideal for
- *   background services, worklets, or apps needing manual lifecycle control.
+ *   {@link createLLMChatSession}, {@link createPaddleOcr},
+ *   {@link createWhisperSpeechToText}) with built-in preprocessing,
+ *   postprocessing, and tokenizers. Ideal for background services, worklets, or
+ *   apps needing manual lifecycle control.
  *
  * - **Domain Utilities (`cv`, `llm`, `speech`, `nlp`, `math`)**:
  *   Domain-specific low-level tools:
  *   - `cv`: Image transformations (resizing, normalization, layout conversions),
- *     bounding box and keypoint helpers.
+ *     bounding box, quadrilateral, and keypoint helpers.
  *   - `llm`: Token-by-token text generation runner, chat template formatting,
  *     and function calling helpers.
  *   - `nlp`: Fast native HuggingFace tokenizers and privacy filter utilities.
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 690aae6322..105276527a 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -2308,6 +2308,9 @@ export const models = {
      * On Android, `VULKAN` is the faster choice.
      */
     PADDLE: {
+      /**
+       * PP-OCRv6 Small multilingual OCR model.
+       */
       PPOCRV6_SMALL: {
         ...PPOCRV6_SMALL_XNNPACK_INT8,
         XNNPACK: PPOCRV6_SMALL_XNNPACK_INT8,

From 737071c5e640e035c4094cfdfb438b494aeeae20 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 13:34:27 +0200
Subject: [PATCH 19/38] docs: refine category taxonomy and add explicit task
 return types

---
 .eslintrc.js                                  | 31 +++++--
 apps/speech/app/text-to-speech/index.tsx      |  9 ++-
 .../react-native-executorch/src/constants.ts  | 38 ++++-----
 .../react-native-executorch/src/core/error.ts | 10 +--
 .../react-native-executorch/src/core/model.ts |  8 +-
 .../src/core/runtime.ts                       |  4 +-
 .../src/core/schema.ts                        | 55 ++++++-------
 .../src/core/tensor.ts                        |  6 +-
 .../src/extensions/cv/image.ts                |  9 ++-
 .../src/extensions/cv/ops/boxes.ts            | 18 ++---
 .../src/extensions/cv/ops/image.ts            | 22 ++---
 .../src/extensions/cv/ops/points.ts           | 10 +--
 .../src/extensions/cv/ops/quad.ts             | 16 ++--
 .../src/extensions/cv/tasks/classification.ts | 68 +++++++++-------
 .../src/extensions/cv/tasks/imageEmbedding.ts | 52 ++++++------
 .../cv/tasks/instanceSegmentation.ts          | 66 ++++++++-------
 .../extensions/cv/tasks/keypointDetection.ts  | 80 +++++++++---------
 .../extensions/cv/tasks/objectDetection.ts    | 62 +++++++-------
 .../src/extensions/cv/tasks/paddleOcr.ts      | 76 +++++++++--------
 .../src/extensions/cv/tasks/preprocessing.ts  | 46 ++++++-----
 .../extensions/cv/tasks/sdxsTextToImage.ts    | 46 ++++++-----
 .../cv/tasks/semanticSegmentation.ts          | 81 ++++++++++---------
 .../src/extensions/cv/tasks/styleTransfer.ts  | 50 +++++++-----
 .../src/extensions/cv/utils/paddleOcrUtils.ts |  4 +-
 .../src/extensions/llm/llmRunner.ts           | 16 ++--
 .../extensions/llm/tasks/llmChatSession.ts    | 12 +--
 .../extensions/llm/utils/chatPreprocessor.ts  | 18 ++---
 .../extensions/llm/utils/tokenizerConfig.ts   |  4 +-
 .../src/extensions/llm/utils/toolCalling.ts   | 10 +--
 .../src/extensions/math.ts                    | 18 ++---
 .../src/extensions/nlp/tasks/privacyFilter.ts | 55 +++++++------
 .../src/extensions/nlp/tasks/textEmbedding.ts | 60 +++++++-------
 .../src/extensions/nlp/tasks/tokenization.ts  | 26 +++++-
 .../src/extensions/nlp/tokenizer.ts           |  4 +-
 .../nlp/utils/privacyFilterUtils.ts           | 13 +--
 .../tasks/fsmnVoiceActivityDetection.ts       | 73 +++++++++--------
 .../speech/tasks/kokoroTextToSpeech.ts        | 56 +++++++------
 .../speech/tasks/supertonicTextToSpeech.ts    | 68 +++++++++-------
 .../speech/tasks/whisperSpeechToText.ts       | 69 +++++++++-------
 .../extensions/speech/utils/kokoroUtils.ts    | 14 ++--
 .../src/extensions/speech/utils/phonemizer.ts |  8 +-
 .../speech/utils/supertonicUtils.ts           | 24 +++---
 .../speech/utils/textPartitioner.ts           |  4 +-
 .../src/extensions/speech/utils/vadUtils.ts   |  4 +-
 .../src/fetcher/fetcher.ts                    |  4 +-
 .../src/fetcher/telemetry.ts                  |  6 +-
 .../src/hooks/useResourceDownload.ts          |  2 +-
 .../src/hooks/useTextToSpeech.ts              | 21 +++--
 packages/react-native-executorch/src/index.ts |  2 +-
 .../react-native-executorch/src/models.ts     |  2 +-
 packages/react-native-executorch/src/utils.ts | 23 ++++--
 51 files changed, 831 insertions(+), 652 deletions(-)

diff --git a/.eslintrc.js b/.eslintrc.js
index e0f544159c..e6c4143afb 100644
--- a/.eslintrc.js
+++ b/.eslintrc.js
@@ -1,13 +1,32 @@
 const path = require('path');
 
 const VALID_CATEGORIES = [
-  'Constants',
-  'Errors',
+  'CV / Constants',
+  'CV / Functions',
+  'CV / Tasks',
+  'CV / Types',
+  'Core / Errors',
+  'Core / Functions',
+  'Core / Schema / Functions',
+  'Core / Schema / Types',
+  'Core / Types',
+  'Fetcher / Functions',
+  'Fetcher / Types',
   'Hooks',
-  'Types',
-  'Typescript API',
-  'Utils',
-  'Utilities - General',
+  'LLM / Functions',
+  'LLM / Tasks',
+  'LLM / Types',
+  'Math / Functions',
+  'Math / Types',
+  'Models',
+  'NLP / Constants',
+  'NLP / Functions',
+  'NLP / Tasks',
+  'NLP / Types',
+  'Speech / Constants',
+  'Speech / Functions',
+  'Speech / Tasks',
+  'Speech / Types',
 ];
 
 const CATEGORY_TAG_MATCH = `^(${VALID_CATEGORIES.join('|')})$`;
diff --git a/apps/speech/app/text-to-speech/index.tsx b/apps/speech/app/text-to-speech/index.tsx
index 7e39eb1e23..ecdb1fcccc 100644
--- a/apps/speech/app/text-to-speech/index.tsx
+++ b/apps/speech/app/text-to-speech/index.tsx
@@ -5,7 +5,8 @@ import {
   models,
   SUPERTONIC_SAMPLE_RATE,
   SUPERTONIC_SUPPORTED_LANGUAGES,
-  constants,
+  SUPERTONIC_DEFAULT_VOICE_NAMES,
+  type SupertonicDefaultVoiceName,
   type SupertonicLanguage,
 } from 'react-native-executorch';
 import { AudioContext, type AudioBufferQueueSourceNode } from 'react-native-audio-api';
@@ -24,9 +25,9 @@ const SAMPLE_TEXT =
   'Each voice style is encoded as a compact embedding that captures the unique timbre, pitch, and speaking patterns of the target speaker. ' +
   'This makes it ideal for accessibility applications, voice assistants, and content creation tools that need high-quality speech synthesis without sending data to external servers.';
 
-const VOICE_OPTIONS = constants.SUPERTONIC_DEFAULT_VOICE_NAMES.map((name) => ({
+const VOICE_OPTIONS = SUPERTONIC_DEFAULT_VOICE_NAMES.map((name) => ({
   label: name,
-  value: name as constants.SupertonicDefaultVoiceName,
+  value: name as SupertonicDefaultVoiceName,
 }));
 
 const LANGUAGE_OPTIONS = SUPERTONIC_SUPPORTED_LANGUAGES.map((lang) => ({
@@ -58,7 +59,7 @@ const MODEL_OPTIONS = [
 function TTSContent() {
   const [text, setText] = useState(SAMPLE_TEXT);
   const [selectedModel, setSelectedModel] = useState<'XNNPACK_FP32' | 'MLX_FP32'>('XNNPACK_FP32');
-  const [selectedVoice, setSelectedVoice] = useState<constants.SupertonicDefaultVoiceName>('F1');
+  const [selectedVoice, setSelectedVoice] = useState<SupertonicDefaultVoiceName>('F1');
   const [selectedLang, setSelectedLang] = useState<SupertonicLanguage>('en');
   const [speed, setSpeed] = useState(1.05);
   const [totalSteps, setTotalSteps] = useState(8);
diff --git a/packages/react-native-executorch/src/constants.ts b/packages/react-native-executorch/src/constants.ts
index d5dd31327c..523361a026 100644
--- a/packages/react-native-executorch/src/constants.ts
+++ b/packages/react-native-executorch/src/constants.ts
@@ -8,7 +8,7 @@
 
 /**
  * ImageNet 1K dataset label array containing the 1000 categories.
- * @category Constants
+ * @category CV / Constants
  */
 export const IMAGENET1K_LABELS = [
   'tench, Tinca tinca',
@@ -1015,7 +1015,7 @@ export const IMAGENET1K_LABELS = [
 
 /**
  * Pascal VOC dataset label array containing the 21 categories.
- * @category Constants
+ * @category CV / Constants
  */
 export const PASCAL_VOC_LABELS = [
   'background',
@@ -1043,7 +1043,7 @@ export const PASCAL_VOC_LABELS = [
 
 /**
  * COCO classes list.
- * @category Constants
+ * @category CV / Constants
  */
 export const COCO_CLASSES = [
   'background',
@@ -1141,25 +1141,25 @@ export const COCO_CLASSES = [
 
 /**
  * Type representing a valid ImageNet 1K label string.
- * @category Types
+ * @category CV / Types
  */
 export type ImageNet1KLabel = (typeof IMAGENET1K_LABELS)[number];
 
 /**
  * Type representing a valid Pascal VOC label string.
- * @category Types
+ * @category CV / Types
  */
 export type PascalVocLabel = (typeof PASCAL_VOC_LABELS)[number];
 
 /**
  * Type representing a valid COCO class string.
- * @category Types
+ * @category CV / Types
  */
 export type CocoClass = (typeof COCO_CLASSES)[number];
 
 /**
  * COCO classes list specifically for YOLO models (80 classes, 0-indexed).
- * @category Constants
+ * @category CV / Constants
  */
 export const COCO_CLASSES_YOLO = [
   'person',
@@ -1246,7 +1246,7 @@ export const COCO_CLASSES_YOLO = [
 
 /**
  * Type representing a valid YOLO COCO class label string.
- * @category Types
+ * @category CV / Types
  */
 export type CocoClassYolo = (typeof COCO_CLASSES_YOLO)[number];
 
@@ -1254,7 +1254,7 @@ export type CocoClassYolo = (typeof COCO_CLASSES_YOLO)[number];
  * ImageNet standard normalization options containing alpha and beta
  * coefficients. Based on the standard ImageNet mean [0.485, 0.456, 0.406] and
  * std [0.229, 0.224, 0.225].
- * @category Constants
+ * @category CV / Constants
  */
 export const IMAGENET_NORM = {
   alpha: [1 / (255.0 * 0.229), 1 / (255.0 * 0.224), 1 / (255.0 * 0.225)],
@@ -1263,7 +1263,7 @@ export const IMAGENET_NORM = {
 
 /**
  * BlazeFace landmarks list.
- * @category Constants
+ * @category CV / Constants
  */
 export const BLAZEFACE_LANDMARKS = [
   'leftEye',
@@ -1276,7 +1276,7 @@ export const BLAZEFACE_LANDMARKS = [
 
 /**
  * COCO human pose landmarks list.
- * @category Constants
+ * @category CV / Constants
  */
 export const COCO_LANDMARKS = [
   'nose',
@@ -1300,19 +1300,19 @@ export const COCO_LANDMARKS = [
 
 /**
  * Type representing a valid BlazeFace landmark string.
- * @category Types
+ * @category CV / Types
  */
 export type BlazeFaceLandmark = (typeof BLAZEFACE_LANDMARKS)[number];
 
 /**
  * Type representing a valid COCO human pose landmark string.
- * @category Types
+ * @category CV / Types
  */
 export type CocoLandmark = (typeof COCO_LANDMARKS)[number];
 
 /**
  * Default Supertonic voice names array.
- * @category Constants
+ * @category Speech / Constants
  */
 // prettier-ignore
 export const SUPERTONIC_DEFAULT_VOICE_NAMES = [
@@ -1322,7 +1322,7 @@ export const SUPERTONIC_DEFAULT_VOICE_NAMES = [
 
 /**
  * Type representing a valid Supertonic default voice name.
- * @category Types
+ * @category Speech / Types
  */
 export type SupertonicDefaultVoiceName = (typeof SUPERTONIC_DEFAULT_VOICE_NAMES)[number];
 
@@ -1336,7 +1336,7 @@ const bioesLabels = <Entity extends string>(entities: readonly Entity[]) =>
 /**
  * Label space for the openai/privacy-filter base model (8 entity types, 33
  * labels).
- * @category Constants
+ * @category NLP / Constants
  */
 export const PRIVACY_FILTER_OPENAI_LABELS = bioesLabels([
   'account_number',
@@ -1352,7 +1352,7 @@ export const PRIVACY_FILTER_OPENAI_LABELS = bioesLabels([
 /**
  * A single BIOES label from the openai/privacy-filter label space (`'O'` or a
  * `B-`/`I-`/`E-`/`S-` prefixed entity).
- * @category Constants
+ * @category NLP / Types
  */
 export type PrivacyFilterOpenaiLabel = (typeof PRIVACY_FILTER_OPENAI_LABELS)[number];
 
@@ -1360,7 +1360,7 @@ export type PrivacyFilterOpenaiLabel = (typeof PRIVACY_FILTER_OPENAI_LABELS)[num
  * Label space for the OpenMed/privacy-filter-nemotron model (55 entity types,
  * 221 labels). Source:
  * https://huggingface.co/OpenMed/privacy-filter-nemotron/blob/main/config.json
- * @category Constants
+ * @category NLP / Constants
  */
 export const PRIVACY_FILTER_NEMOTRON_LABELS = bioesLabels([
   'account_number',
@@ -1423,6 +1423,6 @@ export const PRIVACY_FILTER_NEMOTRON_LABELS = bioesLabels([
 /**
  * A single BIOES label from the OpenMed/privacy-filter-nemotron label space
  * (`'O'` or a `B-`/`I-`/`E-`/`S-` prefixed entity).
- * @category Constants
+ * @category NLP / Types
  */
 export type PrivacyFilterNemotronLabel = (typeof PRIVACY_FILTER_NEMOTRON_LABELS)[number];
diff --git a/packages/react-native-executorch/src/core/error.ts b/packages/react-native-executorch/src/core/error.ts
index 74601124e0..2d97b85a71 100644
--- a/packages/react-native-executorch/src/core/error.ts
+++ b/packages/react-native-executorch/src/core/error.ts
@@ -19,7 +19,7 @@
  * resource, re-create a disposed one). Everything else is a category that
  * exists so crash reporters can group failures, and the detail lives in the
  * message.
- * @category Errors
+ * @category Core / Errors
  */
 export const VALID_ERROR_CODES = [
   'LOAD_FAILED',
@@ -38,7 +38,7 @@ export const VALID_ERROR_CODES = [
  * Machine-readable classification of an {@link RnExecuTorchError}. Branch on
  * this rather than on the message, which is written for humans and can be
  * reworded in any release.
- * @category Errors
+ * @category Core / Errors
  */
 export type RnExecuTorchErrorCode = (typeof VALID_ERROR_CODES)[number];
 
@@ -54,7 +54,7 @@ export type RnExecuTorchErrorCode = (typeof VALID_ERROR_CODES)[number];
  * ```typescript
  * throw RnExecuTorchError('INVALID_ARGUMENT', 'Shape dimensions must be positive');
  * ```
- * @category Errors
+ * @category Core / Errors
  * @typeParam C The specific code, narrowed by {@link isRnExecuTorchError}.
  */
 export type RnExecuTorchError<C extends RnExecuTorchErrorCode = RnExecuTorchErrorCode> = Error & {
@@ -70,7 +70,7 @@ export type RnExecuTorchError<C extends RnExecuTorchErrorCode = RnExecuTorchErro
 
 /**
  * Creates an {@link RnExecuTorchError} carrying a machine-readable code.
- * @category Errors
+ * @category Core / Errors
  * @typeParam C The specific error code.
  * @param code The classification to attach. See {@link RnExecuTorchErrorCode}.
  * @param message A human-readable description of the failure.
@@ -105,7 +105,7 @@ export function RnExecuTorchError<C extends RnExecuTorchErrorCode>(
  *
  * Duck-typed so it holds for errors that crossed a worklet or JSI boundary,
  * where class identity is gone.
- * @category Errors
+ * @category Core / Errors
  * @param err The caught value.
  * @param code When given, also requires the error to carry exactly this code.
  * @returns Whether `err` is an `RnExecuTorchError` (of code `code`, if given).
diff --git a/packages/react-native-executorch/src/core/model.ts b/packages/react-native-executorch/src/core/model.ts
index 867369fc1c..a82bce3e3a 100644
--- a/packages/react-native-executorch/src/core/model.ts
+++ b/packages/react-native-executorch/src/core/model.ts
@@ -15,13 +15,13 @@ declare const modelBrand: unique symbol;
 
 /**
  * A value that can be passed as an input to a model's `execute` method.
- * @category Types
+ * @category Core / Types
  */
 export type ModelInput = Tensor | number | boolean | null;
 
 /**
  * A value returned from a model's `execute` method.
- * @category Types
+ * @category Core / Types
  */
 export type ModelOutput = Tensor | number | boolean | string | null;
 
@@ -34,7 +34,7 @@ export type ModelOutput = Tensor | number | boolean | string | null;
  *
  * Obtain a `Model` instance via the {@link loadModel} function. When the model
  * is no longer needed, call {@link Model.dispose} to release native memory.
- * @category Types
+ * @category Core / Types
  */
 export type Model = {
   /** The local filesystem path of the `.pte` model file. */
@@ -83,7 +83,7 @@ export type Model = {
  * The model is loaded synchronously into native memory. Prefer calling this
  * inside a worklet runtime thread (via {@link wrapAsync}) to avoid blocking the
  * JS thread during compilation.
- * @category Typescript API
+ * @category Core / Functions
  * @param modelPath The absolute local path to the `.pte` model file.
  * @returns The compiled {@link Model} instance, ready for execution.
  * @throws {RnExecuTorchError} Thrown with code `LOAD_FAILED` if the model file
diff --git a/packages/react-native-executorch/src/core/runtime.ts b/packages/react-native-executorch/src/core/runtime.ts
index 6525dee6d6..4fde38e09b 100644
--- a/packages/react-native-executorch/src/core/runtime.ts
+++ b/packages/react-native-executorch/src/core/runtime.ts
@@ -21,7 +21,7 @@ import { isRnExecuTorchError, RnExecuTorchError } from './error';
  * thread, preventing model loading and inference from blocking the UI. Pass it
  * explicitly (or a custom {@link WorkletRuntime}) to {@link wrapAsync} when you
  * need fine-grained control over which thread work executes on.
- * @category Utils
+ * @category Core / Functions
  */
 export const defaultWorkletRuntime = createWorkletRuntime({
   name: 'ExecuTorchDefaultRuntime',
@@ -35,7 +35,7 @@ export const defaultWorkletRuntime = createWorkletRuntime({
  * runtime, awaits the result, and re-throws any error thrown inside the worklet
  * as an {@link RnExecuTorchError}. This keeps heavy native operations (model
  * loading, tensor computation) off the React Native JS thread.
- * @category Utils
+ * @category Core / Functions
  * @typeParam Args The tuple of argument types of `fn`.
  * @typeParam R The return type of `fn`.
  * @param fn A synchronous worklet function to execute on the background
diff --git a/packages/react-native-executorch/src/core/schema.ts b/packages/react-native-executorch/src/core/schema.ts
index a040788991..229147a0ad 100644
--- a/packages/react-native-executorch/src/core/schema.ts
+++ b/packages/react-native-executorch/src/core/schema.ts
@@ -78,7 +78,7 @@ import { RnExecuTorchError } from './error';
 /**
  * Inclusive integer domain of a single dynamic dimension — values from `min`
  * to `max` in increments of `step`.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type Range = { readonly min: number; readonly max: number; readonly step: number };
 
@@ -87,7 +87,7 @@ export type Range = { readonly min: number; readonly max: number; readonly step:
  * - `constant` — exactly `value`.
  * - `range` — any value of a {@link Range}.
  * - `enum` — one of the listed `choices`.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type ConcreteDim =
   | { readonly kind: 'constant'; readonly value: number }
@@ -100,7 +100,7 @@ export type ConcreteDim =
  * validation: `static` symbols bind to constants, `dynamic` symbols to ranges
  * or enums. Reusing a symbol requires every occurrence to bind to the same
  * domain — it does NOT imply any runtime relation between the dimensions.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type SymbolicDim =
   | ConcreteDim
@@ -110,7 +110,7 @@ export type SymbolicDim =
 /**
  * Spec of a tensor parameter: the expected element `dtype` and one
  * dimension spec per axis.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type TensorSpec<Dim extends SymbolicDim> = {
   readonly kind: 'Tensor';
@@ -121,7 +121,7 @@ export type TensorSpec<Dim extends SymbolicDim> = {
 /**
  * The ExecuTorch value-tag that classifies the runtime type of a model input or
  * output slot.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type ExecuTorchTag =
   | 'None'
@@ -138,7 +138,7 @@ export type ExecuTorchTag =
 /**
  * Spec of a single input or output parameter of a method — either a
  * {@link TensorSpec} or a primitive ExecuTorch value tag (`Int`, `Bool`, ...).
- * @category Types
+ * @category Core / Schema / Types
  */
 export type ParamSpec<Dim extends SymbolicDim> =
   | TensorSpec<Dim>
@@ -152,7 +152,7 @@ export type ParamSpec<Dim extends SymbolicDim> =
  * Reference to a single tensor dimension of a method's input or output.
  * `tensorIdx` counts only tensor parameters (skipping primitives), consistent
  * with ExecuTorch's `inputTensorMeta` / `outputTensorMeta` ordering.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type DimRef = {
   readonly paramSide: 'input' | 'output';
@@ -163,7 +163,7 @@ export type DimRef = {
 /**
  * Runtime constraint declaring that all referenced dimensions must be equal
  * to each other in any given execution of the method.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type EqualityConstraint = {
   readonly kind: 'equality';
@@ -174,7 +174,7 @@ export type EqualityConstraint = {
  * Runtime constraint declaring that two dimensions must satisfy
  * `dimLhs = coefficients[0] * dimRhs + coefficients[1]` (integer
  * coefficients) in any given execution of the method.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type LinearConstraint = {
   readonly kind: 'linear';
@@ -187,7 +187,7 @@ export type LinearConstraint = {
  * A requirement on the runtime values of a method's tensor dimensions: the
  * concrete tensors passed to and produced by the method must satisfy it in
  * any given execution. Matched as a declaration during spec validation.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type RuntimeConstraint = LinearConstraint | EqualityConstraint;
 
@@ -198,7 +198,7 @@ export type RuntimeConstraint = LinearConstraint | EqualityConstraint;
 /**
  * Spec of a single model method: the ordered input and output parameter specs
  * and the runtime constraints the method declares over its tensor dimensions.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type MethodSpec<Dim extends SymbolicDim> = {
   inputs: readonly ParamSpec<Dim>[];
@@ -210,7 +210,7 @@ export type MethodSpec<Dim extends SymbolicDim> = {
  * Spec of a whole model, mapping method names to their {@link MethodSpec}.
  * A `SymbolicDim` spec describes allowed models; a `ConcreteDim` spec
  * describes an exported model.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type ModelSpec<Dim extends SymbolicDim> = Record<string, MethodSpec<Dim>>;
 
@@ -222,14 +222,14 @@ export type ModelSpec<Dim extends SymbolicDim> = Record<string, MethodSpec<Dim>>
  * Shape notation accepted by {@link SymbolicTensor}: numbers become
  * {@link ConstantDim}, strings become {@link StaticDim}, and
  * {@link SymbolicDim} values are used as-is.
- * @category Types
+ * @category Core / Schema / Types
  */
 export type SymbolicShape = readonly (number | string | SymbolicDim)[];
 
 /**
  * Creates a static symbolic dimension. Static symbols bind to constant
  * dimensions of the exported spec; repeated uses must bind to the same value.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param symbol The symbol name.
  * @returns The symbolic dimension.
  */
@@ -240,7 +240,7 @@ export const StaticDim = (symbol: string): SymbolicDim => {
 /**
  * Creates a dynamic symbolic dimension. Dynamic symbols bind to range or enum
  * dimensions of the exported spec; repeated uses must bind to the same domain.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param symbol The symbol name.
  * @returns The symbolic dimension.
  */
@@ -250,7 +250,7 @@ export const DynamicDim = (symbol: string): SymbolicDim => {
 
 /**
  * Creates a constant dimension matching exactly `value`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param value The required dimension size.
  * @returns The concrete dimension.
  * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if `value`
@@ -268,7 +268,7 @@ export const ConstantDim = (value: number): ConcreteDim => {
 
 /**
  * Creates an enumerated dimension matching one of `choices`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param choices The allowed dimension sizes.
  * @returns The concrete dimension.
  * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if any
@@ -284,7 +284,7 @@ export const EnumDim = (choices: readonly number[]): ConcreteDim => {
 /**
  * Creates a range dimension matching values from `min` to `max` in increments
  * of `step`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param min The smallest allowed dimension size.
  * @param max The largest allowed dimension size.
  * @param step The increment between allowed sizes. Defaults to 1.
@@ -323,7 +323,7 @@ export const RangeDim = (min: number, max: number, step?: number): ConcreteDim =
 /**
  * Creates a {@link TensorSpec} from a dtype and a {@link SymbolicShape}:
  * numbers become {@link ConstantDim}, strings become {@link StaticDim}.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param dtype The expected element data type.
  * @param shape The per-dimension specs.
  * @returns The tensor spec.
@@ -339,35 +339,35 @@ export const SymbolicTensor = (dtype: DType, shape: SymbolicShape) => {
 
 /**
  * Shorthand for `SymbolicTensor('float32', shape)`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param shape Dimension sizes of the tensor.
  * @returns A {@link SymbolicTensor} with `float32` data type.
  */
 export const f32 = (...shape: SymbolicShape) => SymbolicTensor('float32', shape);
 /**
  * Shorthand for `SymbolicTensor('int64', shape)`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param shape Dimension sizes of the tensor.
  * @returns A {@link SymbolicTensor} with `int64` data type.
  */
 export const i64 = (...shape: SymbolicShape) => SymbolicTensor('int64', shape);
 /**
  * Shorthand for `SymbolicTensor('int32', shape)`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param shape Dimension sizes of the tensor.
  * @returns A {@link SymbolicTensor} with `int32` data type.
  */
 export const i32 = (...shape: SymbolicShape) => SymbolicTensor('int32', shape);
 /**
  * Shorthand for `SymbolicTensor('uint8', shape)`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param shape Dimension sizes of the tensor.
  * @returns A {@link SymbolicTensor} with `uint8` data type.
  */
 export const ui8 = (...shape: SymbolicShape) => SymbolicTensor('uint8', shape);
 /**
  * Shorthand for `SymbolicTensor('bool', shape)`.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param shape Dimension sizes of the tensor.
  * @returns A {@link SymbolicTensor} with `bool` data type.
  */
@@ -375,7 +375,7 @@ export const bool = (...shape: SymbolicShape) => SymbolicTensor('bool', shape);
 
 /**
  * Helper namespace for declaring runtime constraints.
- * @category Typescript API
+ * @category Core / Schema / Functions
  */
 export const constr = {
   /**
@@ -403,7 +403,7 @@ export const constr = {
 /**
  * Constructs a method specification mapping a method name to its parameter
  * specs and runtime constraints.
- * @category Typescript API
+ * @category Core / Schema / Functions
  * @param name The execution method name (e.g., `'forward'`).
  * @param inputs Ordered list of parameter specifications for method inputs.
  * @param outputs Ordered list of parameter specifications for method outputs.
@@ -781,6 +781,7 @@ function validateDimDomains(modelSpec: ModelSpec<SymbolicDim>): void {
 
 /**
  * Result of validating an exported model spec against allowed variants.
+ * @category Core / Schema / Types
  * @typeParam K The variant key type.
  */
 export type SpecMatch<K extends PropertyKey = PropertyKey> = {
@@ -939,7 +940,7 @@ function createSpecMatch<K extends PropertyKey>(
  * accessors.
  * @throws {RnExecuTorchError} With code `SCHEMA_MISMATCH`, describing
  * why every variant failed.
- * @category Typescript API
+ * @category Core / Schema / Functions
  */
 export function validateSpec<const T extends Record<string, ModelSpec<SymbolicDim>>>(
   exportedModelSpec: ModelSpec<ConcreteDim>,
diff --git a/packages/react-native-executorch/src/core/tensor.ts b/packages/react-native-executorch/src/core/tensor.ts
index 5d7649accf..07ef37027f 100644
--- a/packages/react-native-executorch/src/core/tensor.ts
+++ b/packages/react-native-executorch/src/core/tensor.ts
@@ -13,7 +13,7 @@ declare const tensorBrand: unique symbol;
 
 /**
  * Element data type of a {@link Tensor}.
- * @category Types
+ * @category Core / Types
  */
 export type DType = 'float32' | 'uint8' | 'int32' | 'int64' | 'bool';
 
@@ -26,7 +26,7 @@ export type DType = 'float32' | 'uint8' | 'int32' | 'int64' | 'bool';
  * {@link Tensor.dispose} when no longer needed to avoid native memory leaks.
  *
  * Create tensors with the {@link tensor} factory function.
- * @category Types
+ * @category Core / Types
  */
 export type Tensor = {
   /** The element data type of the tensor. */
@@ -125,7 +125,7 @@ export type Tensor = {
  * `src` is omitted the buffer contents are undefined. The returned tensor
  * resides in native C++ memory; call {@link Tensor.dispose} when the tensor is
  * no longer needed.
- * @category Typescript API
+ * @category Core / Functions
  * @param dtype The element data type of the tensor.
  * @param shape An array of dimension sizes (e.g. `[1, 3, 224, 224]`).
  * @param src Optional typed array used to initialize the tensor's data. Its
diff --git a/packages/react-native-executorch/src/extensions/cv/image.ts b/packages/react-native-executorch/src/extensions/cv/image.ts
index 54856b02e4..3dc0d280e8 100644
--- a/packages/react-native-executorch/src/extensions/cv/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/image.ts
@@ -1,12 +1,17 @@
+/**
+ * Core image buffer types and pixel formats for Computer Vision.
+ * @module CV/Image
+ */
+
 /**
  * Supported pixel format layouts for image buffers.
- * @category Types
+ * @category CV / Types
  */
 export type ImageFormat = 'rgb' | 'rgba' | 'bgr' | 'bgra' | 'gray';
 
 /**
  * Represents a raw CPU image buffer in HWC (Height, Width, Channel) layout.
- * @category Types
+ * @category CV / Types
  */
 export type ImageBuffer = {
   readonly data: Uint8Array;
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index 500aa51ea9..f18ab777e6 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -11,7 +11,7 @@ import { scalePoint } from './points';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
- * @category Types
+ * @category CV / Types
  */
 export type BoxMap = Readonly<{
   xyxy: Readonly<{ xmin: number; ymin: number; xmax: number; ymax: number }>;
@@ -21,13 +21,13 @@ export type BoxMap = Readonly<{
 
 /**
  * The formats of bounding boxes.
- * @category Types
+ * @category CV / Types
  */
 export type BoxFormat = keyof BoxMap;
 
 /**
  * Representation of a bounding box under a specific format.
- * @category Types
+ * @category CV / Types
  */
 export type BoundingBox<F extends BoxFormat> = F extends any
   ? { readonly format: F } & BoxMap[F]
@@ -35,7 +35,7 @@ export type BoundingBox<F extends BoxFormat> = F extends any
 
 /**
  * Configuration options for scaling bounding box coordinates.
- * @category Types
+ * @category CV / Types
  */
 export type ScaleBoxOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -49,7 +49,7 @@ export type ScaleBoxOptions = {
 /**
  * Decodes bounding box coordinates from a 4-tuple into a structured BoundingBox
  * object.
- * @category Utils
+ * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param tuple A 4-tuple array containing coordinates.
  * @param format The coordinate format to decode into.
@@ -73,7 +73,7 @@ export function decodeBox<F extends BoxFormat>(
 
 /**
  * Scales bounding box coordinates based on scaling options and resize modes.
- * @category Utils
+ * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param box The original BoundingBox.
  * @param options Options defining dimensions and resize modes.
@@ -139,7 +139,7 @@ export function scaleBox<F extends BoxFormat>(
 
 /**
  * Options for Non-Maximum Suppression (NMS).
- * @category Types
+ * @category CV / Types
  */
 export type NmsOptions = {
   /** How bounding box coordinates are interpreted {@link BoxFormat}. */
@@ -158,7 +158,7 @@ export type NmsOptions = {
 /**
  * Executes Non-Maximum Suppression (NMS) on bounding boxes and confidence
  * scores.
- * @category Utils
+ * @category CV / Functions
  * @param boxes Bounding boxes coordinate tensor. Expected shape `[N, 4]` and
  * data type `float32`.
  * @param scores Bounding boxes confidence scores tensor. Expected shape `[N]`
@@ -198,7 +198,7 @@ export function nms(boxes: Tensor, scores: Tensor, options: NmsOptions): number[
  * Note: This operation does not change the image tensor dimensions (it does not crop
  * the shape). Instead, it copies the elements within the box coordinates from
  * `src` to `dst`, and sets all elements outside the box to `0`.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor to write masked values to.
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index 77a600486e..d79ec64ff2 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -13,7 +13,7 @@ import type { ImageFormat } from '../image';
 
 /**
  * Supported color conversion code presets (similar to OpenCV).
- * @category Types
+ * @category CV / Types
  */
 export type ColorConversionCode =
   | 'RGBA2RGB'
@@ -68,19 +68,19 @@ export const FORMAT_CHANNELS: Record<ImageFormat, number> = {
 
 /**
  * Modes for resizing an image tensor to match target dimensions.
- * @category Types
+ * @category CV / Types
  */
 export type ResizeMode = 'stretch' | 'letterbox' | 'crop';
 
 /**
  * Interpolation algorithms used during image resizing.
- * @category Types
+ * @category CV / Types
  */
 export type InterpolationMethod = 'nearest' | 'area' | 'cubic' | 'lanczos' | 'linear';
 
 /**
  * Configuration options for image resize operations.
- * @category Types
+ * @category CV / Types
  */
 export type ResizeOptions = {
   /** How the image is resized (stretch, letterbox, or crop). */
@@ -93,7 +93,7 @@ export type ResizeOptions = {
 
 /**
  * Configuration options for image tensor normalization.
- * @category Types
+ * @category CV / Types
  */
 export type NormalizeOptions = {
   /**
@@ -112,7 +112,7 @@ export type NormalizeOptions = {
  * Resizes an image tensor from a source dimension to a destination dimension.
  *
  * Supports various {@link ResizeMode} and {@link InterpolationMethod} options.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type (e.g. `uint8`, `float32`).
  * @param dst The pre-allocated destination image tensor to write the resized
@@ -141,7 +141,7 @@ export function resize(src: Tensor, dst: Tensor, options?: ResizeOptions): Tenso
 /**
  * Converts the color space of an image tensor using a specified color
  * conversion code.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor to write the converted
@@ -166,7 +166,7 @@ export function cvtColor(src: Tensor, dst: Tensor, code: ColorConversionCode): T
  * (Channel, Height, Width).
  *
  * Commonly required for PyTorch Edge models which expect channels-first inputs.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor in CHW layout. Expected
@@ -188,7 +188,7 @@ export function toChannelsFirst(src: Tensor, dst: Tensor): Tensor {
  *
  * Useful for post-processing model outputs back into channels-last layouts for
  * rendering or display.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
  * (channels-first). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor in HWC layout. Expected
@@ -210,7 +210,7 @@ export function toChannelsLast(src: Tensor, dst: Tensor): Tensor {
  * Computes: `dst[c,h,w] = src[c,h,w] * alpha[c] + beta[c]`. Can normalize
  * uniformly or channel-wise using array options. The result is cast to `dst`
  * tensor's dtype.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
  * (channels-first). Supports any numeric data type (typically `uint8` or
  * `float32`).
@@ -243,7 +243,7 @@ export function normalize(src: Tensor, dst: Tensor, options?: NormalizeOptions):
  * This operation iterates over each index/class ID in the source tensor, looks
  * up its corresponding RGBA color in the provided colormap palette, and writes
  * it to the destination tensor.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source index/mask image tensor. Expected shape `[H, W, 1]` in
  * HWC layout with data type `int32` containing class indices.
  * @param dst The pre-allocated destination image tensor to write the mapped
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 5685849342..05aac1b9c7 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -7,7 +7,7 @@ import type { ResizeMode } from './image';
 
 /**
  * Represents a 2D coordinate point with x and y values.
- * @category Types
+ * @category CV / Types
  */
 export type Point = {
   readonly x: number;
@@ -16,7 +16,7 @@ export type Point = {
 
 /**
  * Euclidean distance between two points.
- * @category Utils
+ * @category CV / Functions
  * @param a The first point.
  * @param b The second point.
  * @returns The distance between `a` and `b`.
@@ -29,7 +29,7 @@ export function distance(a: Point, b: Point): number {
 /**
  * Linearly interpolates between two points: `t = 0` returns `a`, `t = 1`
  * returns `b`, values in between interpolate along the segment.
- * @category Utils
+ * @category CV / Functions
  * @param a The start point.
  * @param b The end point.
  * @param t The interpolation factor.
@@ -42,7 +42,7 @@ export function interpolatePoint(a: Point, b: Point, t: number): Point {
 
 /**
  * Configuration options for scaling 2D point coordinates.
- * @category Types
+ * @category CV / Types
  */
 export type ScalePointOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -56,7 +56,7 @@ export type ScalePointOptions = {
 /**
  * Helper function to scale a 2D point based on resize mode and resolution
  * changes.
- * @category Utils
+ * @category CV / Functions
  * @param point The original coordinate point to scale.
  * @param options Options detailing the scaling factors and resize mode.
  * See {@link ScalePointOptions}.
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
index 09e6299a1f..833fac8ef0 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -16,14 +16,14 @@ import type { ResizeMode } from './image';
  * Helpers that need them as top-left, top-right, bottom-right, bottom-left say
  * so on their `ordered` parameter; pass the quad through {@link orderQuad}
  * first. Never assume a `Quad` you were handed is already ordered.
- * @category Types
+ * @category CV / Types
  */
 export type Quad = readonly [Point, Point, Point, Point];
 
 /**
  * Computes the axis-aligned bounding box enclosing a set of points, in the
  * requested box format. Returns a zero box for empty input.
- * @category Typescript API
+ * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param points The points to enclose.
  * @param format The coordinate format of the returned box.
@@ -74,7 +74,7 @@ export function boundingBoxOfPoints<F extends BoxFormat>(
  * Reorders a quad's corners into the top-left, top-right, bottom-right,
  * bottom-left order the rest of this module assumes, using their
  * coordinate-sum and coordinate-difference extremes.
- * @category Typescript API
+ * @category CV / Functions
  * @param quad The quad whose corners may be in any order.
  * @returns The same corners, ordered TL, TR, BR, BL.
  */
@@ -99,7 +99,7 @@ export function orderQuad(quad: Quad): Quad {
 /**
  * Computes the width and height (in pixels) of an ordered TL,TR,BR,BL quad, taking
  * the longer of each pair of opposite sides.
- * @category Typescript API
+ * @category CV / Functions
  * @param ordered The quad corners ordered TL, TR, BR, BL.
  * @returns The quad's width and height in pixels.
  */
@@ -113,7 +113,7 @@ export function quadSize(ordered: Quad): { width: number; height: number } {
 
 /**
  * Configuration options for scaling quad coordinates.
- * @category Types
+ * @category CV / Types
  */
 export type ScaleQuadOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -127,7 +127,7 @@ export type ScaleQuadOptions = {
 /**
  * Rescales a quad from one frame to another, clamping the result to the target
  * bounds. The counterpart of {@link scaleBox} for quads.
- * @category Typescript API
+ * @category CV / Functions
  * @param quad The quad, expressed in the `from` frame.
  * @param options Options detailing the scaling factors and resize mode.
  * See {@link ScaleQuadOptions}.
@@ -145,7 +145,7 @@ export function scaleQuad(quad: Quad, options: ScaleQuadOptions): Quad {
 
 /**
  * Options for {@link rectifyQuad}.
- * @category Types
+ * @category CV / Types
  */
 export type RectifyQuadOptions = {
   /** Width in px the rectified content occupies inside the destination canvas. */
@@ -161,7 +161,7 @@ export type RectifyQuadOptions = {
  * `dst`: perspective crop, resize to the canvas height, and pad, in one native
  * pass. An axis-aligned bbox is a 4-corner quad, so pass its corners to
  * rectify a box.
- * @category Typescript API
+ * @category CV / Functions
  * @param src The source image, `uint8` `[H, W, C]`.
  * @param dst The pre-allocated destination canvas, `uint8` `[H', W', C]`, with
  * the same channel count as `src`. Must not alias `src`.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
index ff84894811..b730cbff5b 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
@@ -19,7 +19,7 @@ import { RnExecuTorchError } from '../../../core/error';
 /**
  * Options for configuring an image classifier preprocessor and label
  * vocabulary.
- * @category Types
+ * @category CV / Types
  */
 export type ClassifierOptions<L> = ImagePreprocessorOptions & {
   /** Array of class labels matching the model's output vocabulary. */
@@ -28,7 +28,7 @@ export type ClassifierOptions<L> = ImagePreprocessorOptions & {
 
 /**
  * Model configuration required to instantiate a classifier task runner.
- * @category Types
+ * @category CV / Types
  */
 export type ClassifierModel<L> = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -43,7 +43,7 @@ export type ClassifierModel<L> = {
 
 /**
  * Optional configuration parameters for classification inference.
- * @category Types
+ * @category CV / Types
  */
 export type ClassifyOptions = {
   /**
@@ -55,7 +55,7 @@ export type ClassifyOptions = {
 
 /**
  * Result structure representing a single classification prediction.
- * @category Types
+ * @category CV / Types
  */
 export type Classification<L> = {
   /** Predicted class label. */
@@ -65,33 +65,15 @@ export type Classification<L> = {
 };
 
 /**
- * Creates an image classifier runner for executing local Image Classification
- * models.
- *
- * It validates the model inputs and outputs requirements, asserts that the
- * labels array length matches the model's output vocabulary size, pre-allocates
- * the necessary static execution tensors, sets up an image preprocessor, and
- * registers clean disposal hooks to clear all native memory.
- * @category Typescript API
+ * Image classification task runner.
+ * @category CV / Types
  * @typeParam L The type representing the classification labels.
- * @param config Classifier task configuration containing path and options.
- * See {@link ClassifierModel}.
- * @param runtime Optional worklet runtime thread on which to run the model
- * execution.
- * @returns A promise resolving to an object containing classification and
- * disposal controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
- * `SCHEMA_MISMATCH` if model schema does not match classification spec, or
- * `INVALID_ARGUMENT` if labels length does not match model output classes.
  */
-export async function createClassifier<L>(
-  config: ClassifierModel<L>,
-  runtime?: WorkletRuntime
-): Promise<{
+export type Classifier<L> = {
   /**
    * Releases all allocated native resources.
    */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Performs asynchronous image classification on the given input image.
@@ -104,14 +86,42 @@ export async function createClassifier<L>(
    * negative, `RESOURCE_BUSY` if the model is in use, or
    * `RESOURCE_DISPOSED` if disposed.
    */
-  classify: (input: ImageBuffer, options?: ClassifyOptions) => Promise<Classification<L>[]>;
+  readonly classify: (
+    input: ImageBuffer,
+    options?: ClassifyOptions
+  ) => Promise<Classification<L>[]>;
 
   /**
    * Synchronous version of {@link classify} to be executed directly on the
    * caller or worklet thread.
    */
-  classifyWorklet: (input: ImageBuffer, options?: ClassifyOptions) => Classification<L>[];
-}> {
+  readonly classifyWorklet: (input: ImageBuffer, options?: ClassifyOptions) => Classification<L>[];
+};
+
+/**
+ * Creates an image classifier runner for executing local Image Classification
+ * models.
+ *
+ * It validates the model inputs and outputs requirements, asserts that the
+ * labels array length matches the model's output vocabulary size, pre-allocates
+ * the necessary static execution tensors, sets up an image preprocessor, and
+ * registers clean disposal hooks to clear all native memory.
+ * @category CV / Tasks
+ * @typeParam L The type representing the classification labels.
+ * @param config Classifier task configuration containing path and options.
+ * See {@link ClassifierModel}.
+ * @param runtime Optional worklet runtime thread on which to run the model
+ * execution.
+ * @returns A promise resolving to an object containing classification and
+ * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * `SCHEMA_MISMATCH` if model schema does not match classification spec, or
+ * `INVALID_ARGUMENT` if labels length does not match model output classes.
+ */
+export async function createClassifier<L>(
+  config: ClassifierModel<L>,
+  runtime?: WorkletRuntime
+): Promise<Classifier<L>> {
   const { modelPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
index 7a1d59d20d..7b0a3b7e2f 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
@@ -15,7 +15,7 @@ import { createImagePreprocessor, type ImagePreprocessorOptions } from './prepro
 
 /**
  * Model configuration required to instantiate an image embedder task runner.
- * @category Types
+ * @category CV / Types
  */
 export type ImageEmbedderModel = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -28,6 +28,32 @@ export type ImageEmbedderModel = {
   readonly modelOpts: ImagePreprocessorOptions;
 };
 
+/**
+ * Image embedding task runner.
+ * @category CV / Types
+ */
+export type ImageEmbedder = {
+  /**
+   * Releases all allocated native resources.
+   */
+  readonly dispose: () => void;
+
+  /**
+   * Asynchronously computes the embedding vector for the given input image.
+   * @param input The input image buffer.
+   * @returns A promise resolving to the embedding vector.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
+   */
+  readonly embed: (input: ImageBuffer) => Promise<Float32Array>;
+
+  /**
+   * Synchronous version of {@link embed} to be executed directly on the
+   * caller or worklet thread.
+   */
+  readonly embedWorklet: (input: ImageBuffer) => Float32Array;
+};
+
 /**
  * Creates an image embedder for executing local Image Embedding
  * models (e.g. the image encoder of a CLIP model).
@@ -37,7 +63,7 @@ export type ImageEmbedderModel = {
  * disposal hooks to clear all native memory. Pooling and normalization (if any)
  * are baked into the exported `.pte`; this runner simply preprocesses the image,
  * runs the forward pass, and returns the raw embedding vector.
- * @category Typescript API
+ * @category CV / Tasks
  * @param config Image embedder task configuration containing path and options.
  * See {@link ImageEmbedderModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
@@ -50,27 +76,7 @@ export type ImageEmbedderModel = {
 export async function createImageEmbedder(
   config: ImageEmbedderModel,
   runtime?: WorkletRuntime
-): Promise<{
-  /**
-   * Releases all allocated native resources.
-   */
-  dispose: () => void;
-
-  /**
-   * Asynchronously computes the embedding vector for the given input image.
-   * @param input The input image buffer.
-   * @returns A promise resolving to the embedding vector.
-   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
-   * use, or `RESOURCE_DISPOSED` if disposed.
-   */
-  embed: (input: ImageBuffer) => Promise<Float32Array>;
-
-  /**
-   * Synchronous version of {@link embed} to be executed directly on the
-   * caller or worklet thread.
-   */
-  embedWorklet: (input: ImageBuffer) => Float32Array;
-}> {
+): Promise<ImageEmbedder> {
   const { modelPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index fe2d50196b..5831a30061 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -30,7 +30,7 @@ export type { BoxFormat };
 /**
  * Options for configuring an instance segmenter preprocessor, label
  * vocabulary, and threshold parameters.
- * @category Types
+ * @category CV / Types
  * @typeParam F The format type of the bounding box.
  * @typeParam L The label type.
  */
@@ -54,7 +54,7 @@ export type InstanceSegmenterOptions<F extends BoxFormat, L> = Omit<
 
 /**
  * Model configuration required to instantiate an instance segmenter task runner.
- * @category Types
+ * @category CV / Types
  * @typeParam F The format type of the bounding box.
  * @typeParam L The label type.
  */
@@ -71,7 +71,7 @@ export type InstanceSegmenterModel<F extends BoxFormat, L> = {
 
 /**
  * Optional configuration parameters for instance segmentation inference.
- * @category Types
+ * @category CV / Types
  */
 export type SegmentInstancesOptions = {
   /**
@@ -94,7 +94,7 @@ export type SegmentInstancesOptions = {
 /**
  * Result structure representing a single detected instance with its bounding box,
  * segmentation mask, label, and confidence score.
- * @category Types
+ * @category CV / Types
  * @typeParam F The format type of the bounding box.
  * @typeParam L The label type.
  */
@@ -110,33 +110,16 @@ export type InstanceSegmentationResult<F extends BoxFormat, L> = {
 };
 
 /**
- * Creates an instance segmenter runner for executing local Instance
- * Segmentation models.
- *
- * It validates model input/output tensor shapes and types, pre-allocates
- * execution and auxiliary tensors, sets up an image preprocessor, and returns
- * execution and resource management controls.
- * @category Typescript API
- * @typeParam F The bounding box format type.
+ * Instance segmentation task runner.
+ * @category CV / Types
+ * @typeParam F The format type of the bounding box.
  * @typeParam L The label type.
- * @param config Model configuration containing path and options.
- * See {@link InstanceSegmenterModel}.
- * @param runtime Optional worklet runtime thread on which to run the model
- * execution.
- * @returns A promise resolving to an object containing instance segmentation
- * and disposal controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
- * or `SCHEMA_MISMATCH` if model schema does not match instance segmentation
- * spec.
  */
-export async function createInstanceSegmenter<F extends BoxFormat, L>(
-  config: InstanceSegmenterModel<F, L>,
-  runtime?: WorkletRuntime
-): Promise<{
+export type InstanceSegmenter<F extends BoxFormat, L> = {
   /**
    * Releases all allocated native resources.
    */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Performs asynchronous instance segmentation on the given input image.
@@ -148,7 +131,7 @@ export async function createInstanceSegmenter<F extends BoxFormat, L>(
    * index is out of bounds, `RESOURCE_BUSY` if the model is in use, or
    * `RESOURCE_DISPOSED` if disposed.
    */
-  segmentInstances: (
+  readonly segmentInstances: (
     input: ImageBuffer,
     options?: SegmentInstancesOptions
   ) => Promise<InstanceSegmentationResult<F, L>[]>;
@@ -157,11 +140,36 @@ export async function createInstanceSegmenter<F extends BoxFormat, L>(
    * Synchronous version of {@link segmentInstances} to be executed directly on
    * the caller or worklet thread.
    */
-  segmentInstancesWorklet: (
+  readonly segmentInstancesWorklet: (
     input: ImageBuffer,
     options?: SegmentInstancesOptions
   ) => InstanceSegmentationResult<F, L>[];
-}> {
+};
+
+/**
+ * Creates an instance segmenter runner for executing local Instance
+ * Segmentation models.
+ *
+ * It validates model input/output tensor shapes and types, pre-allocates
+ * execution and auxiliary tensors, sets up an image preprocessor, and returns
+ * execution and resource management controls.
+ * @category CV / Tasks
+ * @typeParam F The bounding box format type.
+ * @typeParam L The label type.
+ * @param config Model configuration containing path and options.
+ * See {@link InstanceSegmenterModel}.
+ * @param runtime Optional worklet runtime thread on which to run the model
+ * execution.
+ * @returns A promise resolving to an object containing instance segmentation
+ * and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match instance segmentation
+ * spec.
+ */
+export async function createInstanceSegmenter<F extends BoxFormat, L>(
+  config: InstanceSegmenterModel<F, L>,
+  runtime?: WorkletRuntime
+): Promise<InstanceSegmenter<F, L>> {
   const { modelPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index b7658b695c..1cc5a340a9 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -22,7 +22,7 @@ export type { BoxFormat };
 
 /**
  * Options for configuring a keypoint detector runner.
- * @category Types
+ * @category CV / Types
  */
 export type KeypointDetectorOptions<F extends BoxFormat, L extends PropertyKey> = Omit<
   ImagePreprocessorOptions,
@@ -42,7 +42,7 @@ export type KeypointDetectorOptions<F extends BoxFormat, L extends PropertyKey>
 
 /**
  * Model configuration required to instantiate a keypoint detector task runner.
- * @category Types
+ * @category CV / Types
  */
 export type KeypointDetectorModel<F extends BoxFormat, L extends PropertyKey> = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -57,7 +57,7 @@ export type KeypointDetectorModel<F extends BoxFormat, L extends PropertyKey> =
 
 /**
  * Optional configuration parameters for keypoint detection inference.
- * @category Types
+ * @category CV / Types
  */
 export type DetectKeypointsOptions = {
   /**
@@ -75,14 +75,14 @@ export type DetectKeypointsOptions = {
 /**
  * Plural landmarks mapped by their names to coordinates and detection
  * confidence.
- * @category Types
+ * @category CV / Types
  */
 export type Landmarks<L extends PropertyKey> = Record<L, Point & { readonly confidence: number }>;
 
 /**
  * Result structure representing a single detected bounding box and its
  * associated landmarks.
- * @category Types
+ * @category CV / Types
  */
 export type KeypointDetection<F extends BoxFormat, L extends PropertyKey> = {
   /** Scaled bounding box coordinates matching the input image resolution. */
@@ -93,6 +93,43 @@ export type KeypointDetection<F extends BoxFormat, L extends PropertyKey> = {
   readonly landmarks: Landmarks<L>;
 };
 
+/**
+ * Keypoint and pose detection task runner.
+ * @category CV / Types
+ * @typeParam F The bounding box format.
+ * @typeParam L The landmark labels type.
+ */
+export type KeypointDetector<F extends BoxFormat, L extends PropertyKey> = {
+  /**
+   * Releases all allocated native resources.
+   */
+  readonly dispose: () => void;
+
+  /**
+   * Performs asynchronous keypoint and bounding box detection on the given
+   * input image.
+   * @param input The input image buffer.
+   * @param options Configuration options for keypoint detection.
+   * See {@link DetectKeypointsOptions}.
+   * @returns A promise resolving to the list of keypoint detections.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
+   */
+  readonly detectKeypoints: (
+    input: ImageBuffer,
+    options?: DetectKeypointsOptions
+  ) => Promise<KeypointDetection<F, L>[]>;
+
+  /**
+   * Synchronous version of {@link detectKeypoints} to be executed directly on
+   * the caller or worklet thread.
+   */
+  readonly detectKeypointsWorklet: (
+    input: ImageBuffer,
+    options?: DetectKeypointsOptions
+  ) => KeypointDetection<F, L>[];
+};
+
 /**
  * Post-processes model outputs by applying Non-Maximum Suppression (NMS) and
  * scaling coordinates.
@@ -173,7 +210,7 @@ function postprocess<F extends BoxFormat, L extends PropertyKey>(
  * It validates model inputs and output shapes (bounding boxes, confidence
  * scores, and landmark coordinates), pre-allocates execution tensors, setups
  * preprocessing, and sets up lifecycle disposals.
- * @category Typescript API
+ * @category CV / Tasks
  * @typeParam F The bounding box format.
  * @typeParam L The landmark labels type.
  * @param config Keypoint task configuration containing path and options.
@@ -188,36 +225,7 @@ function postprocess<F extends BoxFormat, L extends PropertyKey>(
 export async function createKeypointDetector<F extends BoxFormat, L extends PropertyKey>(
   config: KeypointDetectorModel<F, L>,
   runtime?: WorkletRuntime
-): Promise<{
-  /**
-   * Releases all allocated native resources.
-   */
-  dispose: () => void;
-
-  /**
-   * Performs asynchronous keypoint and bounding box detection on the given
-   * input image.
-   * @param input The input image buffer.
-   * @param options Configuration options for keypoint detection.
-   * See {@link DetectKeypointsOptions}.
-   * @returns A promise resolving to the list of keypoint detections.
-   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
-   * use, or `RESOURCE_DISPOSED` if disposed.
-   */
-  detectKeypoints: (
-    input: ImageBuffer,
-    options?: DetectKeypointsOptions
-  ) => Promise<KeypointDetection<F, L>[]>;
-
-  /**
-   * Synchronous version of {@link detectKeypoints} to be executed directly on
-   * the caller or worklet thread.
-   */
-  detectKeypointsWorklet: (
-    input: ImageBuffer,
-    options?: DetectKeypointsOptions
-  ) => KeypointDetection<F, L>[];
-}> {
+): Promise<KeypointDetector<F, L>> {
   const { modelPath, modelOpts } = config;
   const { landmarks } = modelOpts;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index 8136352d8d..0fc81e7132 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -22,7 +22,7 @@ export type { BoxFormat };
 /**
  * Options for configuring an object detector preprocessor, label vocabulary,
  * and detection thresholds.
- * @category Types
+ * @category CV / Types
  */
 export type ObjectDetectorOptions<F extends BoxFormat, L> = Omit<
   ImagePreprocessorOptions,
@@ -42,7 +42,7 @@ export type ObjectDetectorOptions<F extends BoxFormat, L> = Omit<
 
 /**
  * Model configuration required to instantiate an object detector task runner.
- * @category Types
+ * @category CV / Types
  */
 export type ObjectDetectorModel<F extends BoxFormat, L> = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -57,7 +57,7 @@ export type ObjectDetectorModel<F extends BoxFormat, L> = {
 
 /**
  * Optional configuration parameters for object detection inference.
- * @category Types
+ * @category CV / Types
  */
 export type DetectObjectsOptions = {
   /**
@@ -74,7 +74,7 @@ export type DetectObjectsOptions = {
 
 /**
  * Result structure representing a single object detection prediction.
- * @category Types
+ * @category CV / Types
  */
 export type ObjectDetection<F extends BoxFormat, L> = {
   /** Scaled bounding box coordinates matching the input image resolution. */
@@ -86,32 +86,16 @@ export type ObjectDetection<F extends BoxFormat, L> = {
 };
 
 /**
- * Creates an object detector runner for executing local Object Detection
- * models.
- *
- * It validates the model inputs and outputs requirements, pre-allocates the
- * necessary static execution tensors (boxes, scores, classes), sets up an image
- * preprocessor, and registers clean disposal hooks to clear all native memory.
- * @category Typescript API
+ * Object detection task runner.
+ * @category CV / Types
  * @typeParam F The bounding box format.
  * @typeParam L The type representing the class labels.
- * @param config Object detector task configuration containing path and options.
- * See {@link ObjectDetectorModel}.
- * @param runtime Optional worklet runtime thread on which to run the model
- * execution.
- * @returns A promise resolving to an object containing object detection and
- * disposal controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
- * or `SCHEMA_MISMATCH` if model schema does not match object detection spec.
  */
-export async function createObjectDetector<F extends BoxFormat, L>(
-  config: ObjectDetectorModel<F, L>,
-  runtime?: WorkletRuntime
-): Promise<{
+export type ObjectDetector<F extends BoxFormat, L> = {
   /**
    * Releases all allocated native resources.
    */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Asynchronously performs object detection on the input image.
@@ -123,7 +107,7 @@ export async function createObjectDetector<F extends BoxFormat, L>(
    * index is out of bounds, `RESOURCE_BUSY` if the model is in use, or
    * `RESOURCE_DISPOSED` if disposed.
    */
-  detectObjects: (
+  readonly detectObjects: (
     input: ImageBuffer,
     options?: DetectObjectsOptions
   ) => Promise<ObjectDetection<F, L>[]>;
@@ -132,11 +116,35 @@ export async function createObjectDetector<F extends BoxFormat, L>(
    * Synchronous version of {@link detectObjects} to be executed directly on the
    * caller or worklet thread.
    */
-  detectObjectsWorklet: (
+  readonly detectObjectsWorklet: (
     input: ImageBuffer,
     options?: DetectObjectsOptions
   ) => ObjectDetection<F, L>[];
-}> {
+};
+
+/**
+ * Creates an object detector runner for executing local Object Detection
+ * models.
+ *
+ * It validates the model inputs and outputs requirements, pre-allocates the
+ * necessary static execution tensors (boxes, scores, classes), sets up an image
+ * preprocessor, and registers clean disposal hooks to clear all native memory.
+ * @category CV / Tasks
+ * @typeParam F The bounding box format.
+ * @typeParam L The type representing the class labels.
+ * @param config Object detector task configuration containing path and options.
+ * See {@link ObjectDetectorModel}.
+ * @param runtime Optional worklet runtime thread on which to run the model
+ * execution.
+ * @returns A promise resolving to an object containing object detection and
+ * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match object detection spec.
+ */
+export async function createObjectDetector<F extends BoxFormat, L>(
+  config: ObjectDetectorModel<F, L>,
+  runtime?: WorkletRuntime
+): Promise<ObjectDetector<F, L>> {
   const { modelPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 366b882bd2..5b96c4b8fa 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -44,7 +44,7 @@ import { createImagePreprocessor } from './preprocessing';
 
 /**
  * A single recognized text region.
- * @category Types
+ * @category CV / Types
  */
 export type OcrDetection = {
   /** The recognized text. */
@@ -61,7 +61,7 @@ export type OcrDetection = {
 
 /**
  * Options for the PP-OCRv6 pipeline.
- * @category Types
+ * @category CV / Types
  */
 export type PaddleOcrModelOptions = {
   /**
@@ -73,7 +73,7 @@ export type PaddleOcrModelOptions = {
 
 /**
  * Optional configuration parameters for optical character recognition inference.
- * @category Types
+ * @category CV / Types
  */
 export type RecognizeCharactersOptions = {
   /**
@@ -86,7 +86,7 @@ export type RecognizeCharactersOptions = {
 /**
  * Model configuration for the PP-OCRv6 pipeline: one fused detect/recognize PTE,
  * the charset published beside it, and the run options.
- * @category Types
+ * @category CV / Types
  */
 export type PaddleOcrModel = {
   /** The fused detect/recognize PTE. Resolved to a local path by the fetcher. */
@@ -101,6 +101,38 @@ export type PaddleOcrModel = {
   readonly modelOpts: PaddleOcrModelOptions;
 };
 
+/**
+ * PP-OCRv6 optical character recognition task runner.
+ * @category CV / Types
+ */
+export type PaddleOcr = {
+  /**
+   * Releases all allocated native resources.
+   */
+  readonly dispose: () => void;
+
+  /**
+   * Detects and recognizes every text line in the given image.
+   * @param input The input image buffer.
+   * @param options Per-call overrides. See {@link RecognizeCharactersOptions}.
+   * @returns A promise resolving to the recognized lines in reading order
+   * (leftmost column top to bottom, then the next column).
+   */
+  readonly recognizeCharacters: (
+    input: ImageBuffer,
+    options?: RecognizeCharactersOptions
+  ) => Promise<OcrDetection[]>;
+
+  /**
+   * Synchronous version of {@link recognizeCharacters} to be executed directly
+   * on the caller or worklet thread.
+   */
+  readonly recognizeCharactersWorklet: (
+    input: ImageBuffer,
+    options?: RecognizeCharactersOptions
+  ) => OcrDetection[];
+};
+
 // Fixed by the export: the detector was trained on ImageNet-normalized RGB and
 // the recognizer on (x/255 - 0.5)/0.5 over a gray-padded canvas.
 const DETECTOR_PREPROCESSOR_OPTS = {
@@ -354,9 +386,11 @@ function greedyCtcDecode(
  * Creates the PP-OCRv6 runner: one pass detects text quads on the whole page,
  * warps each to the recognizer canvas and reads it, returning the lines in
  * reading order.
- * @category Typescript API
- * @param config Model path, charset path, and run options. See {@link PaddleOcrModel}.
- * @param runtime Optional worklet runtime thread.
+ * @category CV / Tasks
+ * @param config PaddleOCR task configuration containing the model, charset,
+ * and detection thresholds. See {@link PaddleOcrModel}.
+ * @param runtime Optional worklet runtime thread on which to run detection and
+ * recognition.
  * @returns A promise resolving to an object containing recognition and disposal
  * controls.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model or charset
@@ -367,33 +401,7 @@ function greedyCtcDecode(
 export async function createPaddleOcr(
   config: PaddleOcrModel,
   runtime?: WorkletRuntime
-): Promise<{
-  /**
-   * Releases all allocated native resources.
-   */
-  dispose: () => void;
-
-  /**
-   * Detects and recognizes every text line in the given image.
-   * @param input The input image buffer.
-   * @param options Per-call overrides. See {@link RecognizeCharactersOptions}.
-   * @returns A promise resolving to the recognized lines in reading order
-   * (leftmost column top to bottom, then the next column).
-   */
-  recognizeCharacters: (
-    input: ImageBuffer,
-    options?: RecognizeCharactersOptions
-  ) => Promise<OcrDetection[]>;
-
-  /**
-   * Synchronous version of {@link recognizeCharacters} to be executed directly
-   * on the caller or worklet thread.
-   */
-  recognizeCharactersWorklet: (
-    input: ImageBuffer,
-    options?: RecognizeCharactersOptions
-  ) => OcrDetection[];
-}> {
+): Promise<PaddleOcr> {
   const { modelPath, charsetPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
index b3b7314620..b0b2cbb39e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
@@ -21,7 +21,7 @@ import { RnExecuTorchError } from '../../../core/error';
 
 /**
  * Options for configuring the image preprocessor pipeline.
- * @category Types
+ * @category CV / Types
  */
 export type ImagePreprocessorOptions = {
   /** How the input image is resized to match the model's expected dimensions. */
@@ -34,6 +34,29 @@ export type ImagePreprocessorOptions = {
   readonly padValue?: number;
 };
 
+/**
+ * Image preprocessor runner for transforming image buffers into model input tensors.
+ * @category CV / Types
+ */
+export type ImagePreprocessor = {
+  /**
+   * Preprocesses the input image by resizing, converting color space, changing
+   * format layout, and normalizing values, copying the output directly to the
+   * pre-allocated output tensor.
+   *
+   * Note: The returned tensor is managed by the preprocessor; consumers do not
+   * need to dispose of it manually.
+   * @param input The input image buffer to preprocess.
+   * @returns A reference to the output tensor containing preprocessed float32
+   * data of shape `[3, H, W]` (or `[1, 3, H, W]`) and data type `float32`.
+   */
+  readonly process: (input: ImageBuffer) => Tensor;
+  /**
+   * Releases all allocated native resources.
+   */
+  readonly dispose: () => void;
+};
+
 /**
  * Creates a reusable image preprocessor pipeline.
  *
@@ -41,7 +64,7 @@ export type ImagePreprocessorOptions = {
  * normalize, and copy raw image buffers into target tensors matching model
  * input shapes. All intermediate scratch tensors are pre-allocated and safely
  * disposed of when calling `dispose()`.
- * @category Typescript API
+ * @category CV / Functions
  * @param options Normalization scaling coefficients, interpolation algorithms, and
  * resize modes.
  * See {@link ImagePreprocessorOptions}.
@@ -55,24 +78,7 @@ export type ImagePreprocessorOptions = {
 export function createImagePreprocessor(
   options: ImagePreprocessorOptions,
   outputShape: number[]
-): {
-  /**
-   * Preprocesses the input image by resizing, converting color space, changing
-   * format layout, and normalizing values, copying the output directly to the
-   * pre-allocated output tensor.
-   *
-   * Note: The returned tensor is managed by the preprocessor; consumers do not
-   * need to dispose of it manually.
-   * @param input The input image buffer to preprocess.
-   * @returns A reference to the output tensor containing preprocessed float32
-   * data of shape `[3, H, W]` (or `[1, 3, H, W]`) and data type `float32`.
-   */
-  process: (input: ImageBuffer) => Tensor;
-  /**
-   * Releases all allocated native resources.
-   */
-  dispose: () => void;
-} {
+): ImagePreprocessor {
   'worklet';
   const numRgbChannels = 3;
   const isRank3 = outputShape.length === 3 && outputShape[0] === numRgbChannels;
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
index 379696d5ad..2e45714ee6 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
@@ -41,7 +41,7 @@ const NOISE_COEFF = -14.579279;
 
 /**
  * Model configuration required to instantiate the SDXS text-to-image runner.
- * @category Types
+ * @category CV / Types
  */
 export type SdxsTextToImageModel = {
   /** Local path to the model `.pte`. */
@@ -51,24 +51,12 @@ export type SdxsTextToImageModel = {
 };
 
 /**
- * Creates an SDXS text-to-image runner.
- *
- * It validates the exported method schemas, pre-allocates the static execution
- * tensors, and registers disposal hooks that release all native memory.
- * @category Typescript API
- * @param config SDXS pipeline configuration containing the model and tokenizer paths.
- * See {@link SdxsTextToImageModel}.
- * @param runtime Optional worklet runtime thread on which to run generation.
- * @returns A promise resolving to an object with generation and disposal controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model or tokenizer
- * fails to load, or `SCHEMA_MISMATCH` if model schema does not match SDXS spec.
+ * SDXS single-step text-to-image generation task runner.
+ * @category CV / Types
  */
-export async function createSdxsTextToImage(
-  config: SdxsTextToImageModel,
-  runtime?: WorkletRuntime
-): Promise<{
+export type SdxsTextToImage = {
   /** Releases all allocated native resources. */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Generates an image from a text prompt.
@@ -79,14 +67,32 @@ export async function createSdxsTextToImage(
    * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
    * use, or `RESOURCE_DISPOSED` if disposed.
    */
-  generate: (prompt: string, seed?: number) => Promise<ImageBuffer>;
+  readonly generate: (prompt: string, seed?: number) => Promise<ImageBuffer>;
 
   /**
    * Synchronous version of {@link generate} to be executed directly on the
    * caller or worklet thread.
    */
-  generateWorklet: (prompt: string, seed?: number) => ImageBuffer;
-}> {
+  readonly generateWorklet: (prompt: string, seed?: number) => ImageBuffer;
+};
+
+/**
+ * Creates an SDXS text-to-image runner.
+ *
+ * It validates the exported method schemas, pre-allocates the static execution
+ * tensors, and registers disposal hooks that release all native memory.
+ * @category CV / Tasks
+ * @param config SDXS pipeline configuration containing the model and tokenizer paths.
+ * See {@link SdxsTextToImageModel}.
+ * @param runtime Optional worklet runtime thread on which to run generation.
+ * @returns A promise resolving to an object with generation and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model or tokenizer
+ * fails to load, or `SCHEMA_MISMATCH` if model schema does not match SDXS spec.
+ */
+export async function createSdxsTextToImage(
+  config: SdxsTextToImageModel,
+  runtime?: WorkletRuntime
+): Promise<SdxsTextToImage> {
   const { modelPath, tokenizerPath } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
   const tokenizer = await wrapAsync(loadTokenizer, runtime)(tokenizerPath);
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
index 1c88186016..928518df82 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
@@ -27,7 +27,7 @@ import { RnExecuTorchError } from '../../../core/error';
 /**
  * Options for configuring a semantic segmenter preprocessor and label
  * vocabulary.
- * @category Types
+ * @category CV / Types
  */
 export type SemanticSegmenterOptions<L> = Omit<ImagePreprocessorOptions, 'resizeMode'> & {
   /** Resize mode for input images. Must be `'stretch'`. */
@@ -40,7 +40,7 @@ export type SemanticSegmenterOptions<L> = Omit<ImagePreprocessorOptions, 'resize
 
 /**
  * Model configuration required to instantiate a segmenter task runner.
- * @category Types
+ * @category CV / Types
  */
 export type SemanticSegmenterModel<L> = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -55,13 +55,13 @@ export type SemanticSegmenterModel<L> = {
 
 /**
  * Maps each class label to its assigned RGBA color.
- * @category Types
+ * @category CV / Types
  */
 export type ColorMap<L extends PropertyKey> = Record<L, [number, number, number, number]>;
 
 /**
  * Result structure representing the output of a semantic segmentation task.
- * @category Types
+ * @category CV / Types
  */
 export type SemanticSegmentationResult<L extends PropertyKey> = {
   /** Generated output RGBA image buffer containing the colored segmentation mask. */
@@ -70,42 +70,16 @@ export type SemanticSegmentationResult<L extends PropertyKey> = {
   readonly colormap?: ColorMap<L>;
 };
 
-function hslToRgb(h: number, s: number, l: number): [number, number, number] {
-  s /= 100;
-  l /= 100;
-  const k = (n: number) => (n + h / 30) % 12;
-  const a = s * Math.min(l, 1 - l);
-  const f = (n: number) => l - a * Math.max(-1, Math.min(k(n) - 3, 9 - k(n), 1));
-  return [Math.round(255 * f(0)), Math.round(255 * f(8)), Math.round(255 * f(4))];
-}
-
 /**
- * Creates a semantic segmenter runner for executing local Semantic Segmentation
- * models.
- *
- * It validates the model inputs and outputs, asserts that the labels array
- * length matches the model's output vocabulary size, pre-allocates the
- * necessary static execution tensors, sets up an image preprocessor, and
- * registers clean disposal hooks to clear all native memory.
- * @category Typescript API
+ * Semantic segmentation task runner.
+ * @category CV / Types
  * @typeParam L The type representing the segmentation labels.
- * @param config Segmenter task configuration containing path and options.
- * See {@link SemanticSegmenterModel}.
- * @param runtime Optional worklet runtime thread environment context.
- * @returns A promise resolving to an object containing segmentation and
- * disposal controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
- * `SCHEMA_MISMATCH` if model schema does not match segmentation spec, or
- * `INVALID_ARGUMENT` if labels length does not match model output classes.
  */
-export async function createSemanticSegmenter<L extends PropertyKey = string>(
-  config: SemanticSegmenterModel<L>,
-  runtime?: WorkletRuntime
-): Promise<{
+export type SemanticSegmenter<L extends PropertyKey = string> = {
   /**
    * Releases all allocated native resources.
    */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Runs semantic segmentation asynchronously.
@@ -130,7 +104,7 @@ export async function createSemanticSegmenter<L extends PropertyKey = string>(
    * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
    * use, or `RESOURCE_DISPOSED` if disposed.
    */
-  segment: (
+  readonly segment: (
     input: ImageBuffer,
     colormap?: Partial<ColorMap<L>>
   ) => Promise<SemanticSegmentationResult<L>>;
@@ -139,11 +113,44 @@ export async function createSemanticSegmenter<L extends PropertyKey = string>(
    * Synchronous version of {@link segment} to be executed directly on the
    * caller or worklet thread.
    */
-  segmentWorklet: (
+  readonly segmentWorklet: (
     input: ImageBuffer,
     colormap?: Partial<ColorMap<L>>
   ) => SemanticSegmentationResult<L>;
-}> {
+};
+
+function hslToRgb(h: number, s: number, l: number): [number, number, number] {
+  s /= 100;
+  l /= 100;
+  const k = (n: number) => (n + h / 30) % 12;
+  const a = s * Math.min(l, 1 - l);
+  const f = (n: number) => l - a * Math.max(-1, Math.min(k(n) - 3, 9 - k(n), 1));
+  return [Math.round(255 * f(0)), Math.round(255 * f(8)), Math.round(255 * f(4))];
+}
+
+/**
+ * Creates a semantic segmenter runner for executing local Semantic Segmentation
+ * models.
+ *
+ * It validates the model inputs and outputs, asserts that the labels array
+ * length matches the model's output vocabulary size, pre-allocates the
+ * necessary static execution tensors, sets up an image preprocessor, and
+ * registers clean disposal hooks to clear all native memory.
+ * @category CV / Tasks
+ * @typeParam L The type representing the segmentation labels.
+ * @param config Segmenter task configuration containing path and options.
+ * See {@link SemanticSegmenterModel}.
+ * @param runtime Optional worklet runtime thread environment context.
+ * @returns A promise resolving to an object containing segmentation and
+ * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * `SCHEMA_MISMATCH` if model schema does not match segmentation spec, or
+ * `INVALID_ARGUMENT` if labels length does not match model output classes.
+ */
+export async function createSemanticSegmenter<L extends PropertyKey = string>(
+  config: SemanticSegmenterModel<L>,
+  runtime?: WorkletRuntime
+): Promise<SemanticSegmenter<L>> {
   const { modelPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
index 80642340c3..9a63fe031e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
@@ -24,7 +24,7 @@ import {
 
 /**
  * Options for configuring the style transfer preprocessor and postprocessor.
- * @category Types
+ * @category CV / Types
  */
 export type StyleTransferOptions = Omit<ImagePreprocessorOptions, 'resizeMode'> & {
   /** Resize mode for input images. Must be `'stretch'`. */
@@ -37,7 +37,7 @@ export type StyleTransferOptions = Omit<ImagePreprocessorOptions, 'resizeMode'>
 
 /**
  * Model configuration required to instantiate a style transfer task runner.
- * @category Types
+ * @category CV / Types
  */
 export type StyleTransferModel = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -51,27 +51,14 @@ export type StyleTransferModel = {
 };
 
 /**
- * Creates an image style transfer runner for executing local style transfer models.
- *
- * It validates the model inputs and outputs requirements, pre-allocates
- * the necessary static execution tensors, sets up an image preprocessor, and
- * registers clean disposal hooks to clear all native memory.
- * @category Typescript API
- * @param config Style transfer task configuration containing path and options.
- * See {@link StyleTransferModel}.
- * @param runtime Optional worklet runtime thread on which to run the model execution.
- * @returns A promise resolving to an object containing style transfer and disposal controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
- * or `SCHEMA_MISMATCH` if model schema does not match style transfer spec.
+ * Image style transfer task runner.
+ * @category CV / Types
  */
-export async function createStyleTransfer(
-  config: StyleTransferModel,
-  runtime?: WorkletRuntime
-): Promise<{
+export type StyleTransfer = {
   /**
    * Releases all allocated native resources.
    */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Performs asynchronous image style transfer on the given input image.
@@ -80,14 +67,33 @@ export async function createStyleTransfer(
    * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
    * use, or `RESOURCE_DISPOSED` if disposed.
    */
-  transferStyle: (input: ImageBuffer) => Promise<ImageBuffer>;
+  readonly transferStyle: (input: ImageBuffer) => Promise<ImageBuffer>;
 
   /**
    * Synchronous version of {@link transferStyle} to be executed directly on the
    * caller or worklet thread.
    */
-  transferStyleWorklet: (input: ImageBuffer) => ImageBuffer;
-}> {
+  readonly transferStyleWorklet: (input: ImageBuffer) => ImageBuffer;
+};
+
+/**
+ * Creates an image style transfer runner for executing local style transfer models.
+ *
+ * It validates the model inputs and outputs requirements, pre-allocates
+ * the necessary static execution tensors, sets up an image preprocessor, and
+ * registers clean disposal hooks to clear all native memory.
+ * @category CV / Tasks
+ * @param config Style transfer task configuration containing path and options.
+ * See {@link StyleTransferModel}.
+ * @param runtime Optional worklet runtime thread on which to run the model execution.
+ * @returns A promise resolving to an object containing style transfer and disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
+ * or `SCHEMA_MISMATCH` if model schema does not match style transfer spec.
+ */
+export async function createStyleTransfer(
+  config: StyleTransferModel,
+  runtime?: WorkletRuntime
+): Promise<StyleTransfer> {
   const { modelPath, modelOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
diff --git a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
index 4c6aa97434..af53fec8fd 100644
--- a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
@@ -31,7 +31,7 @@ function quadsFromFlat(flat: ArrayLike<number>): Quad[] {
 
 /**
  * Thresholds for {@link extractDbnetTextQuads}.
- * @category Types
+ * @category CV / Types
  */
 export type DbnetDecodeOptions = {
   /** Binarization threshold on the probability map. */
@@ -50,7 +50,7 @@ export type DbnetDecodeOptions = {
  * Decodes a DBNet probability map into oriented text quads: binarizes the map,
  * traces contours, scores each candidate by its mean probability and unclips the
  * survivors back to their unshrunk size.
- * @category Typescript API
+ * @category CV / Functions
  * @param probabilityMap The `detect` output, shape `[1, 1, H, W]`, post-sigmoid.
  * @param options Decode thresholds. See {@link DbnetDecodeOptions}.
  * @returns The decoded quads, in detector-input pixel space and arbitrary order.
diff --git a/packages/react-native-executorch/src/extensions/llm/llmRunner.ts b/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
index 4a93503af4..7681c31d3f 100644
--- a/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
+++ b/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
@@ -11,7 +11,7 @@ declare const llmRunnerBrand: unique symbol;
 /**
  * Configuration options for LLM text generation.
  * @experimental This API is experimental and might change in future releases.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMGenerationConfig = {
   /** Whether to echo the prompt in the generated output. */
@@ -27,7 +27,7 @@ export type LLMGenerationConfig = {
 /**
  * Execution and performance statistics for a generation call.
  * @experimental This API is experimental and might change in future releases.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMGenerationStats = {
   /** Number of tokens in the input prompt. */
@@ -51,7 +51,7 @@ export type LLMGenerationStats = {
 /**
  * Low-level non-text media input tensor payloads.
  * @experimental This API is experimental and might change in future releases.
- * @category Types
+ * @category LLM / Types
  */
 export type MediaInput =
   | { readonly kind: 'image'; readonly image: Tensor }
@@ -60,21 +60,21 @@ export type MediaInput =
 /**
  * Supported non-text input modality keys (e.g. `'image'`, `'audio'`).
  * @experimental This API is experimental and might change in future releases.
- * @category Types
+ * @category LLM / Types
  */
 export type Modality = MediaInput['kind'];
 
 /**
  * Text or interleaved multimodal prompt input for a low-level LLM runner.
  * @experimental This API is experimental and might change in future releases.
- * @category Types
+ * @category LLM / Types
  */
 export type Prompt = string | readonly (string | MediaInput)[];
 
 /**
  * Current KV cache state and capacity metrics for an LLM runner.
  * @experimental This API is experimental and might change in future releases.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMKVCacheState = {
   /** Current token position index / number of occupied tokens in the KV cache. */
@@ -92,7 +92,7 @@ export type LLMKVCacheState = {
  * @experimental This API is experimental and might change in future releases. It
  * relies on experimental ExecuTorch runtime extensions and injected member-pointer
  * accessors to manage KV cache state that may evolve across releases.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMRunner = {
   /** Path to the local model file. */
@@ -155,7 +155,7 @@ export type LLMRunner = {
  * @experimental This API is experimental and might change in future releases. It
  * relies on experimental ExecuTorch runtime extensions and injected member-pointer
  * accessors to manage KV cache state that may evolve across releases.
- * @category Typescript API
+ * @category LLM / Functions
  * @param modelPath Path to the local `.pte` model file.
  * @param tokenizerPath Path to the local tokenizer configuration file (e.g. `tokenizer.json`).
  * @param modalities List of supported input non-text modalities (e.g.
diff --git a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
index 9f8acf4e14..c1a1d7b24a 100644
--- a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
+++ b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
@@ -48,7 +48,7 @@ export type {
 
 /**
  * Model configuration required to instantiate an LLM chat session.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMModel = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -65,7 +65,7 @@ export type LLMModel = {
 
 /**
  * Configuration options for tool calling in an LLM chat session.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMToolOpts = {
   /** Tool definitions available to the model. */
@@ -78,7 +78,7 @@ export type LLMToolOpts = {
 
 /**
  * Options for configuring an LLM chat session.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMChatSessionOptions = {
   /** Default generation configuration options. */
@@ -98,7 +98,7 @@ export type LLMChatSessionOptions = {
 
 /**
  * Result returned by an LLM chat turn.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMChatTurnResult = {
   /** The messages added to history during this chat turn. */
@@ -115,7 +115,7 @@ export type LLMChatTurnResult = {
 
 /**
  * Handle to an active LLM chat session.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMChatSession = {
   /**
@@ -186,7 +186,7 @@ const DEFAULT_MAX_TURNS = 5;
 
 /**
  * Instantiates an LLM chat session using background thread execution.
- * @category Typescript API
+ * @category LLM / Tasks
  * @param config Model configuration containing model, tokenizer, and tokenizer config paths.
  * @param options Custom generation, tool calling, and state options.
  * @param runtime The worklet runtime thread to run native generation on.
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
index 531d92f1ab..e98545c753 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
@@ -19,7 +19,7 @@ import type { ToolDefinition, ToolCall } from './toolCalling';
 
 /**
  * High-level media payload input for chat turns.
- * @category Types
+ * @category LLM / Types
  */
 export type ChatMediaInput =
   | { readonly kind: 'image'; readonly image: ImageBuffer }
@@ -27,13 +27,13 @@ export type ChatMediaInput =
 
 /**
  * Interleaved text and media content for a chat turn.
- * @category Types
+ * @category LLM / Types
  */
 export type ChatMessageContent = string | readonly (string | ChatMediaInput)[];
 
 /**
  * Conversation turn representing system, user, assistant, or tool execution messages.
- * @category Types
+ * @category LLM / Types
  */
 export type ChatMessage = Readonly<
   | { role: 'system' | 'user'; content: ChatMessageContent }
@@ -43,7 +43,7 @@ export type ChatMessage = Readonly<
 
 /**
  * Image preprocessing and sentinel token config for vision-language LLMs.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMImagePreprocessorConfig = {
   /** Sentinel token delimiters inserted into Jinja prompts. */
@@ -56,7 +56,7 @@ export type LLMImagePreprocessorConfig = {
 
 /**
  * Audio preprocessing and sentinel token config for audio-language LLMs.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMAudioPreprocessorConfig = {
   /** Sentinel token delimiters inserted into Jinja prompts. */
@@ -65,7 +65,7 @@ export type LLMAudioPreprocessorConfig = {
 
 /**
  * Preprocessor configuration for media modalities.
- * @category Types
+ * @category LLM / Types
  */
 export type LLMMediaPreprocessorConfig = {
   /** Image preprocessing configuration for vision-language models. */
@@ -76,7 +76,7 @@ export type LLMMediaPreprocessorConfig = {
 
 /**
  * Options for instantiating a ChatPreprocessor.
- * @category Types
+ * @category LLM / Types
  */
 export type ChatPreprocessorConfig = {
   /** Jinja chat template string for prompt rendering. */
@@ -92,7 +92,7 @@ export type ChatPreprocessorConfig = {
 /**
  * Chat preprocessor object for Jinja template rendering, prompt diffing, and
  * media tensor preprocessing.
- * @category Types
+ * @category LLM / Types
  */
 export type ChatPreprocessor = {
   /**
@@ -296,7 +296,7 @@ function prepareStringMessages(
 /**
  * Handles Jinja template formatting and media tensor preprocessing for chat
  * turns.
- * @category Typescript API
+ * @category LLM / Functions
  * @param config Preprocessor configuration including template and preprocessor
  * settings.
  * @returns A ChatPreprocessor object containing process, render, buildPrompt,
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
index 6a5cba0c86..4b80a51b35 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
@@ -8,7 +8,7 @@ import { RnExecuTorchError } from '../../../core/error';
 
 /**
  * Model chat template configuration resolved from tokenizer config file.
- * @category Types
+ * @category LLM / Types
  */
 export type TokenizerChatConfig = {
   /** Jinja chat template string for prompt rendering. */
@@ -27,7 +27,7 @@ function resolveToken(token: unknown): string | undefined {
 
 /**
  * Parses raw JSON configuration from `tokenizer_config.json` into a normalized format.
- * @category Utils
+ * @category LLM / Functions
  * @param config Raw JSON object from tokenizer_config.json.
  * @returns A parsed TokenizerChatConfig object.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if `chat_template` is not
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
index 0ab0d4c00b..1ceec65d07 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
@@ -8,7 +8,7 @@ import type { ChatMessageContent } from './chatPreprocessor';
 
 /**
  * JSON Schema definition for tool parameter inputs.
- * @category Types
+ * @category LLM / Types
  */
 export type ToolParameters = {
   /** JSON Schema type, typically `'object'`. */
@@ -22,7 +22,7 @@ export type ToolParameters = {
 
 /**
  * Declaration of a tool available to the model.
- * @category Types
+ * @category LLM / Types
  */
 export type ToolDefinition<Args extends Record<string, unknown> = Record<string, unknown>> = {
   /** Tool type discriminator, typically `'function'`. */
@@ -45,7 +45,7 @@ export type ToolDefinition<Args extends Record<string, unknown> = Record<string,
 
 /**
  * Function call object embedded inside assistant tool_calls.
- * @category Types
+ * @category LLM / Types
  */
 export type ToolCall = {
   /** Optional tool call identifier for matching responses. */
@@ -62,7 +62,7 @@ export type ToolCall = {
 
 /**
  * Structured tool parsing result containing detected tool calls and remaining text.
- * @category Types
+ * @category LLM / Types
  */
 export type ToolParserResult = {
   /** Detected tool calls extracted from the model output. */
@@ -74,6 +74,6 @@ export type ToolParserResult = {
 /**
  * Function signature for parsing tool calls from model output text.
  * Returns undefined if no tool call was detected.
- * @category Types
+ * @category LLM / Types
  */
 export type ToolParser = (text: string) => ToolParserResult | undefined;
diff --git a/packages/react-native-executorch/src/extensions/math.ts b/packages/react-native-executorch/src/extensions/math.ts
index cd27320a19..fe10f7bac4 100644
--- a/packages/react-native-executorch/src/extensions/math.ts
+++ b/packages/react-native-executorch/src/extensions/math.ts
@@ -9,7 +9,7 @@ import { RnExecuTorchError } from '../core/error';
 
 /**
  * Configuration options for normal distribution random number generation.
- * @category Types
+ * @category Math / Types
  */
 export type RandomNormalOptions = {
   /** The mean of the distribution. */
@@ -26,7 +26,7 @@ export type RandomNormalOptions = {
 /**
  * Computes the element-wise sigmoid activation on a float32 source tensor and
  * writes the result to a destination tensor.
- * @category Typescript API
+ * @category Math / Functions
  * @param src The input source tensor. Expected shape `[d1, ..., dn]` with data
  * type `float32`.
  * @param dst The pre-allocated destination tensor to write the result to.
@@ -45,7 +45,7 @@ export function sigmoid(src: Tensor, dst: Tensor): Tensor {
 /**
  * Computes the softmax activation along a specified axis on a float32 source
  * tensor and writes the result to a destination tensor.
- * @category Typescript API
+ * @category Math / Functions
  * @param src The input source tensor. Expected shape `[d1, ..., dn]` with data
  * type `float32`.
  * @param dst The pre-allocated destination tensor to write the result to.
@@ -66,7 +66,7 @@ export function softmax(src: Tensor, dst: Tensor, axis: number = -1): Tensor {
 /**
  * Computes the indices of the maximum values along a specified axis on a
  * float32 source tensor and writes the result to an int32 destination tensor.
- * @category Typescript API
+ * @category Math / Functions
  * @param src The input source tensor. Expected shape `[d1, ..., dk, ..., dn]`
  * with data type `float32`.
  * @param dst The pre-allocated destination tensor to write the indices to.
@@ -90,7 +90,7 @@ export function argmax(src: Tensor, dst: Tensor, axis: number = -1): Tensor {
  * given by an int32 index tensor. Pairs with {@link argmax}, whose output has
  * exactly the shape this expects, so `argmax` then `gather` yields the maximum
  * values alongside their indices.
- * @category Typescript API
+ * @category Math / Functions
  * @param src The input float32 source tensor. Shape [d1,...,dk,...,dn].
  * @param indices The int32 index tensor, one index per lane. Shape
  * [d1,...,1,...,dn].
@@ -110,7 +110,7 @@ export function gather(src: Tensor, indices: Tensor, dst: Tensor, axis: number =
 /**
  * Applies the element-wise threshold step function on a float32 source tensor and
  * writes the result to a destination tensor.
- * @category Typescript API
+ * @category Math / Functions
  * @param src The input source tensor. Expected shape `[d1, ..., dn]` with data
  * type `float32`.
  * @param dst The pre-allocated destination tensor to write the result to.
@@ -132,7 +132,7 @@ export function threshold(src: Tensor, dst: Tensor, thresholdVal: number): Tenso
  * Creates a mulberry32 pseudo-random generator producing uniform values in
  * `[0, 1)`. Unlike `Math.random` it accepts a seed, so a fixed seed yields a
  * reproducible sequence.
- * @category Typescript API
+ * @category Math / Functions
  * @param seed The 32-bit integer seed for the generator.
  * @returns A function returning the next uniform pseudo-random number in `[0, 1)`.
  */
@@ -155,7 +155,7 @@ export function mulberry32(seed: number): () => number {
 /**
  * Draws normally distributed values using the Box–Muller transform, seeded via
  * {@link mulberry32} so a fixed `seed` reproduces the same sequence.
- * @category Typescript API
+ * @category Math / Functions
  * @param size The number of values to draw.
  * @param options Distribution parameters. When options or any individual
  * properties are omitted, defaults to `mean: 0`, `std: 1`, and a random seed.
@@ -186,7 +186,7 @@ export function randomNormal(size: number, options?: RandomNormalOptions): Float
  * kind — the equivalent of PyTorch's `repeat_interleave` over a 1-D input.
  *
  * Non-positive repeat counts drop their element.
- * @category Typescript API
+ * @category Math / Functions
  * @typeParam T The array kind of `values` (any typed array or a plain array),
  * preserved in the result.
  * @param values The values to repeat.
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
index 72598ff805..2586840ccc 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
@@ -28,7 +28,7 @@ import {
 /**
  * Options describing a privacy filter model's label space and decoding
  * behavior.
- * @category Types
+ * @category NLP / Types
  * @typeParam Label The model's BIOES label space, defined alongside the model
  * in the `models` registry.
  */
@@ -53,7 +53,7 @@ export type PrivacyFilterOptions<Label extends string = string> = {
 
 /**
  * Model configuration required to instantiate a privacy filter task runner.
- * @category Types
+ * @category NLP / Types
  * @typeParam Label The model's BIOES label space.
  */
 export type PrivacyFilterModel<Label extends string = string> = {
@@ -71,6 +71,33 @@ export type PrivacyFilterModel<Label extends string = string> = {
 export { piiSegments };
 export type { PiiEntity, PiiEntityType, PiiSegment, ViterbiBiases };
 
+/**
+ * Privacy filter task runner for detecting PII entities in text.
+ * @category NLP / Types
+ * @typeParam Label The model's BIOES label space.
+ */
+export type PrivacyFilter<Label extends string = string> = {
+  /**
+   * Releases all allocated native resources.
+   */
+  readonly dispose: () => void;
+
+  /**
+   * Asynchronously detects PII entity spans in the given text.
+   * @param input The text string to scan for PII.
+   * @returns A promise resolving to the detected entity spans, in order.
+   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
+   * use, or `RESOURCE_DISPOSED` if disposed.
+   */
+  readonly detectPii: (input: string) => Promise<PiiEntity<PiiEntityType<Label>>[]>;
+
+  /**
+   * Synchronous version of {@link detectPii} to be executed directly on the
+   * caller or worklet thread.
+   */
+  readonly detectPiiWorklet: (input: string) => PiiEntity<PiiEntityType<Label>>[];
+};
+
 /**
  * Creates a privacy filter runner that detects personally identifiable
  * information (PII) spans in text.
@@ -85,7 +112,7 @@ export type { PiiEntity, PiiEntityType, PiiSegment, ViterbiBiases };
  * windows with 50% overlap and never truncated; predictions near a window's
  * edges are discarded in favor of the neighboring window's more centered
  * context.
- * @category Typescript API
+ * @category NLP / Tasks
  * @typeParam Label The model's BIOES label space.
  * @param config Privacy filter task configuration containing the model and
  * tokenizer paths plus the label space options. See {@link PrivacyFilterModel}.
@@ -101,27 +128,7 @@ export type { PiiEntity, PiiEntityType, PiiSegment, ViterbiBiases };
 export async function createPrivacyFilter<Label extends string>(
   config: PrivacyFilterModel<Label>,
   runtime?: WorkletRuntime
-): Promise<{
-  /**
-   * Releases all allocated native resources.
-   */
-  dispose: () => void;
-
-  /**
-   * Asynchronously detects PII entity spans in the given text.
-   * @param input The text string to scan for PII.
-   * @returns A promise resolving to the detected entity spans, in order.
-   * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
-   * use, or `RESOURCE_DISPOSED` if disposed.
-   */
-  detectPii: (input: string) => Promise<PiiEntity<PiiEntityType<Label>>[]>;
-
-  /**
-   * Synchronous version of {@link detectPii} to be executed directly on the
-   * caller or worklet thread.
-   */
-  detectPiiWorklet: (input: string) => PiiEntity<PiiEntityType<Label>>[];
-}> {
+): Promise<PrivacyFilter<Label>> {
   const { modelPath, tokenizerPath, modelOpts } = config;
   const { labelNames } = modelOpts;
 
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
index 5ede1a72cd..8d0e4ca08c 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
@@ -15,7 +15,7 @@ import { RnExecuTorchError } from '../../../core/error';
 
 /**
  * Model configuration required to instantiate a text embedder task runner.
- * @category Types
+ * @category NLP / Types
  */
 export type TextEmbedderModel = {
   /** Local path or remote URL of the `.pte` model file. */
@@ -26,6 +26,36 @@ export type TextEmbedderModel = {
   readonly defaultPrompt?: string;
 };
 
+/**
+ * Text embedding task runner.
+ * @category NLP / Types
+ */
+export type TextEmbedder = {
+  /**
+   * Releases all allocated native resources.
+   */
+  readonly dispose: () => void;
+
+  /**
+   * Asynchronously computes the embedding vector for the given input text.
+   * Inputs longer than the model's maximum sequence length are truncated.
+   * @param input The input text to embed.
+   * @param prompt Optional prompt prefix overriding the model's configured
+   * `defaultPrompt` for this call.
+   * @returns A promise resolving to the embedding vector.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the input text
+   * tokenizes to zero tokens, `RESOURCE_BUSY` if the model is in use, or
+   * `RESOURCE_DISPOSED` if disposed.
+   */
+  readonly embed: (input: string, prompt?: string) => Promise<Float32Array>;
+
+  /**
+   * Synchronous version of {@link embed} to be executed directly on the
+   * caller or worklet thread.
+   */
+  readonly embedWorklet: (input: string, prompt?: string) => Float32Array;
+};
+
 /**
  * Creates a text embedder for executing local Text Embedding models
  * (e.g. sentence-transformers like all-MiniLM-L6-v2).
@@ -37,7 +67,7 @@ export type TextEmbedderModel = {
  * model's maximum sequence length; the attention mask is all ones. Pooling and
  * normalization are baked into the exported `.pte`; this runner runs the forward
  * pass and returns the raw embedding vector.
- * @category Typescript API
+ * @category NLP / Tasks
  * @param config Text embedder task configuration containing the model and
  * tokenizer paths. See {@link TextEmbedderModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
@@ -51,31 +81,7 @@ export type TextEmbedderModel = {
 export async function createTextEmbedder(
   config: TextEmbedderModel,
   runtime?: WorkletRuntime
-): Promise<{
-  /**
-   * Releases all allocated native resources.
-   */
-  dispose: () => void;
-
-  /**
-   * Asynchronously computes the embedding vector for the given input text.
-   * Inputs longer than the model's maximum sequence length are truncated.
-   * @param input The input text to embed.
-   * @param prompt Optional prompt prefix overriding the model's configured
-   * `defaultPrompt` for this call.
-   * @returns A promise resolving to the embedding vector.
-   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the input text
-   * tokenizes to zero tokens, `RESOURCE_BUSY` if the model is in use, or
-   * `RESOURCE_DISPOSED` if disposed.
-   */
-  embed: (input: string, prompt?: string) => Promise<Float32Array>;
-
-  /**
-   * Synchronous version of {@link embed} to be executed directly on the
-   * caller or worklet thread.
-   */
-  embedWorklet: (input: string, prompt?: string) => Float32Array;
-}> {
+): Promise<TextEmbedder> {
   const { modelPath, tokenizerPath, defaultPrompt } = config;
   const [model, tokenizer] = await Promise.all([
     wrapAsync(loadModel, runtime)(modelPath),
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
index bda219d259..9b0942ba82 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
@@ -8,16 +8,38 @@ import type { WorkletRuntime } from 'react-native-worklets';
 import { wrapAsync } from '../../../core/runtime';
 import { loadTokenizer } from '../tokenizer';
 
+/**
+ * Asynchronous tokenizer task runner wrapping native HuggingFace tokenizer operations.
+ * @category NLP / Types
+ */
+export type Tokenizer = {
+  /** Encodes a string into token ids asynchronously. */
+  readonly encode: (text: string) => Promise<Int32Array>;
+  /** Decodes token ids back into a string asynchronously. */
+  readonly decode: (tokens: Int32Array, skipSpecialTokens?: boolean) => Promise<string>;
+  /** Returns the total vocabulary size. */
+  readonly getVocabSize: () => number;
+  /** Converts a numeric token id to its string piece representation. */
+  readonly idToToken: (id: number) => string;
+  /** Converts a string piece token to its numeric token id. */
+  readonly tokenToId: (token: string) => number;
+  /** Releases the native tokenizer resources. */
+  readonly dispose: () => void;
+};
+
 /**
  * Loads a tokenizer and exposes its operations with lifetime management for the
  * `useTokenizer` hook.
- * @category Typescript API
+ * @category NLP / Tasks
  * @param tokenizerPath Absolute local path to a `tokenizer.json` file.
  * @param runtime Optional worklet runtime thread to run the tokenizer on.
  * @returns A promise resolving to the tokenizer operations and a `dispose`
  * handle that releases the native tokenizer.
  */
-export async function createTokenizer(tokenizerPath: string, runtime?: WorkletRuntime) {
+export async function createTokenizer(
+  tokenizerPath: string,
+  runtime?: WorkletRuntime
+): Promise<Tokenizer> {
   const tokenizer = await wrapAsync(loadTokenizer, runtime)(tokenizerPath);
   const dispose = () => tokenizer.dispose();
 
diff --git a/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts b/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
index c4d4d5bacd..2690a9f0f2 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
@@ -10,7 +10,7 @@ declare const tokenizerBrand: unique symbol;
 /**
  * A native HuggingFace-compatible tokenizer instance backed by a JSI host
  * object. All methods are synchronous and worklet-compatible.
- * @category Types
+ * @category NLP / Types
  */
 export type Tokenizer = {
   /** Absolute local file path of the loaded `tokenizer.json`. */
@@ -83,7 +83,7 @@ export type Tokenizer = {
 
 /**
  * Loads a HuggingFace tokenizer from a local `tokenizer.json` file.
- * @category Typescript API
+ * @category NLP / Functions
  * @param tokenizerPath Absolute local path to a `tokenizer.json` file.
  * @returns The loaded native {@link Tokenizer} instance.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the tokenizer file
diff --git a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
index 1065aa4cd8..5a20ce2885 100644
--- a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
@@ -22,7 +22,7 @@ const NEG_INF = Number.NEGATIVE_INFINITY;
  * transition is taken. Positive values encourage the transition; negative
  * values discourage it. Every field defaults to `0` (a neutral, validity-only
  * Viterbi).
- * @category Types
+ * @category NLP / Types
  */
 export interface ViterbiBiases {
   /** `O -> O` (background persistence). Higher = stay in background more, fewer false positives. */
@@ -44,14 +44,14 @@ export interface ViterbiBiases {
  * `B-`/`I-`/`E-`/`S-` prefix (`'O'` and any unprefixed label contribute
  * nothing). Lets a model's concrete label list narrow {@link PiiEntity.label}
  * to just that model's entity types instead of an arbitrary string.
- * @category Types
+ * @category NLP / Types
  */
 export type PiiEntityType<Label extends string> =
   Label extends `${'B' | 'I' | 'E' | 'S'}-${infer Entity}` ? Entity : never;
 
 /**
  * A single detected PII entity span.
- * @category Types
+ * @category NLP / Types
  * @typeParam Label The entity type, narrowed to a specific model's entity types
  * when known (see {@link PiiEntityType}), or `string` for an arbitrary model.
  */
@@ -90,7 +90,7 @@ const CLASS_S = 4;
  * background state, the best span-closing (`E-`/`S-`) state, or, per entity,
  * the best of that entity's `B-`/`I-` states — which is what lets the decode
  * step run in `O(numLabels + numEntities)` instead of `O(numLabels^2)`.
- * @category Types
+ * @category NLP / Types
  */
 export interface Grammar {
   /** Total number of labels, i.e. `1 + 4 * numEntities`. */
@@ -391,7 +391,7 @@ function isSpanOpener(labelId: number, labelNames: readonly string[]): boolean {
 
 /**
  * A contiguous run of same-entity tokens, before text decoding.
- * @category Types
+ * @category NLP / Types
  */
 export interface TokenSpan {
   readonly start: number;
@@ -473,7 +473,7 @@ export function computeCharOffsets(tokenizer: Tokenizer, ids: Int32Array): Uint3
  * One run in a segmentation of source text: either a plain (uncovered) run or
  * a labeled entity slice. Adjacent segments cover the input without gaps or
  * overlaps; concatenating every `text` reproduces the input.
- * @category Types
+ * @category NLP / Types
  * @typeParam Label The entity label type; narrows to a specific model's
  * entity types when known, or `string` for an arbitrary model.
  */
@@ -490,6 +490,7 @@ export type PiiSegment<Label extends string = string> =
  * index into it) and must not overlap; the pipeline never emits overlapping
  * spans, so the typical case is passing `detectPii`'s result straight through.
  * Overlapping inputs are resolved first-wins by start position.
+ * @category NLP / Functions
  * @param text The original input the entities were detected in.
  * @param entities Detected entity spans, in any order.
  * @returns A flat list of plain and entity segments covering `text`.
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
index 099651ebde..1e0f57812f 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
@@ -13,7 +13,7 @@ import { extractFrames } from '../utils/vadUtils';
 
 /**
  * Sample rate (Hz) the FSMN-VAD model expects its input waveform to be at.
- * @category Constants
+ * @category Speech / Constants
  */
 export const FSMN_VAD_SAMPLE_RATE_HZ = 16000;
 
@@ -37,7 +37,7 @@ const DEFAULT_DETECTION_MARGIN_MS = 100;
 /**
  * Tunable thresholds controlling how per-frame speech probabilities are turned
  * into speech {@link Segment}s.
- * @category Types
+ * @category Speech / Types
  */
 export type VadOptions = {
   /** Minimum speech probability (0-1) for a frame to count as speech. */
@@ -57,7 +57,7 @@ export type VadOptions = {
 
 /**
  * Model configuration required to instantiate an FSMN-VAD task runner.
- * @category Types
+ * @category Speech / Types
  */
 export type FsmnVadModel = {
   /** Local path or remote URL of the `.pte` model. */
@@ -72,7 +72,7 @@ export type FsmnVadModel = {
 
 /**
  * A detected speech region, with start and end expressed in seconds.
- * @category Types
+ * @category Speech / Types
  */
 export type Segment = {
   /** Start time of the speech segment in seconds. */
@@ -84,7 +84,7 @@ export type Segment = {
 /**
  * Options controlling live detection via `detectVoiceOnStream`. Extends the
  * per-call detection thresholds ({@link VadOptions}).
- * @category Types
+ * @category Speech / Types
  */
 export type VadStreamOptions = VadOptions & {
   /**
@@ -96,7 +96,7 @@ export type VadStreamOptions = VadOptions & {
 
 /**
  * A speech-activity transition reported by `detectVoiceOnStream`.
- * @category Types
+ * @category Speech / Types
  */
 export type VadEvent = 'speechStart' | 'speechEnd';
 
@@ -174,31 +174,14 @@ function postprocess(scores: Float32Array, options: Required<VadOptions>): Segme
 }
 
 /**
- * Creates a Voice Activity Detection runner for the FSMN-VAD model.
- *
- * It loads the model, validates its input/output signature and registers a
- * disposal hook. The whole pipeline — feature extraction, chunked inference and
- * segment postprocessing — runs in TypeScript on top of the core `model.execute`
- * primitive.
- * @category Typescript API
- * @param config VAD task configuration containing the model path and default
- * options. See {@link FsmnVadModel}.
- * @param runtime Optional worklet runtime thread on which to run the model
- * execution.
- * @returns A promise resolving to an object containing detection and disposal
- * controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model fails to
- * load, or `SCHEMA_MISMATCH` if the model schema does not match the VAD
- * specification.
+ * Voice activity detection task runner using the FSMN-VAD model.
+ * @category Speech / Types
  */
-export async function createFsmnVoiceActivityDetector(
-  config: FsmnVadModel,
-  runtime?: WorkletRuntime
-): Promise<{
+export type FsmnVoiceActivityDetector = {
   /**
    * Releases all allocated native resources.
    */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Asynchronously detects speech segments within a mono waveform sampled at
@@ -210,13 +193,13 @@ export async function createFsmnVoiceActivityDetector(
    * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
    * use, or `RESOURCE_DISPOSED` if disposed.
    */
-  detectVoice: (waveform: Float32Array, options?: VadOptions) => Promise<Segment[]>;
+  readonly detectVoice: (waveform: Float32Array, options?: VadOptions) => Promise<Segment[]>;
 
   /**
    * Synchronous version of {@link detectVoice} to be executed directly on the caller
    * or worklet thread.
    */
-  detectVoiceWorklet: (waveform: Float32Array, options?: VadOptions) => Segment[];
+  readonly detectVoiceWorklet: (waveform: Float32Array, options?: VadOptions) => Segment[];
 
   /**
    * Appends a live audio chunk to a bounded rolling window, runs detection over
@@ -231,13 +214,39 @@ export async function createFsmnVoiceActivityDetector(
    * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
    * use, or `RESOURCE_DISPOSED` if disposed.
    */
-  detectVoiceOnStream: (chunk: Float32Array, options?: VadStreamOptions) => VadEvent | undefined;
+  readonly detectVoiceOnStream: (
+    chunk: Float32Array,
+    options?: VadStreamOptions
+  ) => VadEvent | undefined;
 
   /**
    * Clears the rolling window and speaking state used by {@link detectVoiceOnStream}.
    */
-  resetStream: () => void;
-}> {
+  readonly resetStream: () => void;
+};
+
+/**
+ * Creates a Voice Activity Detection runner for the FSMN-VAD model.
+ *
+ * It loads the model, validates its input/output signature and registers a
+ * disposal hook. The whole pipeline — feature extraction, chunked inference and
+ * segment postprocessing — runs in TypeScript on top of the core `model.execute`
+ * primitive.
+ * @category Speech / Tasks
+ * @param config VAD task configuration containing the model path and default
+ * options. See {@link FsmnVadModel}.
+ * @param runtime Optional worklet runtime thread on which to run the model
+ * execution.
+ * @returns A promise resolving to an object containing detection and disposal
+ * controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model fails to
+ * load, or `SCHEMA_MISMATCH` if the model schema does not match the VAD
+ * specification.
+ */
+export async function createFsmnVoiceActivityDetector(
+  config: FsmnVadModel,
+  runtime?: WorkletRuntime
+): Promise<FsmnVoiceActivityDetector> {
   const { modelPath, defaultOptions } = config;
 
   const model = await wrapAsync(loadModel, runtime)(modelPath);
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
index 60882c3085..1b117c1d33 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
@@ -35,7 +35,7 @@ import {
 
 /**
  * Kokoro audio sampling rate in Hz (24000).
- * @category Constants
+ * @category Speech / Constants
  */
 export const KOKORO_SAMPLE_RATE = 24000;
 
@@ -54,7 +54,7 @@ const LETTER_PATTERN = /\p{L}/u;
 
 /**
  * Model configuration required to instantiate the Kokoro Text-to-Speech pipeline.
- * @category Types
+ * @category Speech / Types
  * @typeParam K Voice keys record constraint (strictly inferred from voices keys).
  */
 export type KokoroTtsModel<K extends PropertyKey> = {
@@ -75,7 +75,7 @@ export type KokoroTtsModel<K extends PropertyKey> = {
 
 /**
  * Per-call execution options for Kokoro Text-to-Speech synthesis.
- * @category Types
+ * @category Speech / Types
  * @typeParam K Voice keys record constraint.
  */
 export type KokoroTtsOptions<K extends PropertyKey> = {
@@ -91,7 +91,7 @@ export type KokoroTtsOptions<K extends PropertyKey> = {
 
 /**
  * Audio output chunk yielded by the {@link createKokoroTextToSpeech} generator.
- * @category Types
+ * @category Speech / Types
  */
 export type KokoroTtsChunk = {
   /** Float32 PCM audio samples for this chunk, normalized in `[-1, 1]`. */
@@ -106,13 +106,40 @@ export type KokoroTtsChunk = {
   readonly totalChunks: number;
 };
 
+/**
+ * Kokoro text-to-speech task runner.
+ * @category Speech / Types
+ * @typeParam K Voice keys record constraint.
+ */
+export type KokoroTextToSpeech<K extends PropertyKey = string> = {
+  /** Releases the allocated native models, phonemizer and execution tensors. */
+  readonly dispose: () => void;
+
+  /**
+   * Streams synthesized audio chunks as an async generator as each text chunk finishes.
+   * @param text Input text (or IPA phonemes) to synthesize into speech.
+   * @param options Per-call execution options. See {@link KokoroTtsOptions}.
+   * @returns An AsyncGenerator yielding {@link KokoroTtsChunk} audio buffers.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the voice is unknown
+   * or speed is out of range, `RESOURCE_BUSY` if the model is in use, or
+   * `RESOURCE_DISPOSED` if disposed.
+   */
+  readonly synthesize: (
+    text: string,
+    options: KokoroTtsOptions<K>
+  ) => AsyncGenerator<KokoroTtsChunk>;
+
+  /** Cancels any in-flight synthesis started by {@link synthesize}. */
+  readonly synthesizeStop: () => void;
+};
+
 /**
  * Creates a Kokoro Text-to-Speech pipeline.
  *
  * It validates both sub-model method schemas, builds the grapheme-to-phoneme
  * pipeline, pre-parses the voice files into memory, and registers disposal
  * hooks to release all native resources.
- * @category Typescript API
+ * @category Speech / Tasks
  * @typeParam K Voice keys record constraint.
  * @param config Kokoro TTS pipeline configuration containing model and asset paths.
  * See {@link KokoroTtsModel}.
@@ -125,24 +152,7 @@ export type KokoroTtsChunk = {
 export async function createKokoroTextToSpeech<K extends PropertyKey>(
   config: KokoroTtsModel<K>,
   runtime?: WorkletRuntime
-): Promise<{
-  /** Releases the allocated native models, phonemizer and execution tensors. */
-  dispose: () => void;
-
-  /**
-   * Streams synthesized audio chunks as an async generator as each text chunk finishes.
-   * @param text Input text (or IPA phonemes) to synthesize into speech.
-   * @param options Per-call execution options. See {@link KokoroTtsOptions}.
-   * @returns An AsyncGenerator yielding {@link KokoroTtsChunk} audio buffers.
-   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the voice is unknown
-   * or speed is out of range, `RESOURCE_BUSY` if the model is in use, or
-   * `RESOURCE_DISPOSED` if disposed.
-   */
-  synthesize: (text: string, options: KokoroTtsOptions<K>) => AsyncGenerator<KokoroTtsChunk>;
-
-  /** Cancels any in-flight synthesis started by {@link synthesize}. */
-  synthesizeStop: () => void;
-}> {
+): Promise<KokoroTextToSpeech<K>> {
   const load = wrapAsync(loadModel, runtime);
   const [durationPredictor, synthesizer] = await Promise.all([
     load(config.modelPaths.durationPredictor),
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
index 9bfe32f3a1..68d14c5244 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
@@ -28,7 +28,7 @@ export { SUPERTONIC_SUPPORTED_LANGUAGES, type SupertonicVoiceStyle, type Superto
 
 /**
  * Supertonic audio sampling rate in Hz (44100).
- * @category Constants
+ * @category Speech / Constants
  */
 export const SUPERTONIC_SAMPLE_RATE = 44100;
 
@@ -61,9 +61,11 @@ function getDefaultMaxChunkLength(lang?: SupertonicLanguage): number {
 }
 
 /**
- * Model configuration required to instantiate the Supertonic Text-to-Speech pipeline.
- * @category Types
- * @typeParam K Voice style keys record constraint (strictly inferred from voiceStyles keys).
+ * Model configuration required to instantiate the Supertonic Text-to-Speech
+ * pipeline.
+ * @category Speech / Types
+ * @typeParam K Voice style keys record constraint (strictly inferred from
+ * voiceStyles keys).
  */
 export type SupertonicTtsModel<K extends PropertyKey> = {
   /** Discriminates this config from the other Text-to-Speech pipelines. */
@@ -87,7 +89,7 @@ export type SupertonicTtsModel<K extends PropertyKey> = {
 
 /**
  * Per-call execution options for Supertonic Text-to-Speech synthesis.
- * @category Types
+ * @category Speech / Types
  * @typeParam K Voice style keys record constraint.
  */
 export type SupertonicTtsOptions<K extends PropertyKey> = {
@@ -108,7 +110,7 @@ export type SupertonicTtsOptions<K extends PropertyKey> = {
 
 /**
  * Audio output chunk yielded by the {@link createSupertonicTextToSpeech} generator.
- * @category Types
+ * @category Speech / Types
  */
 export type SupertonicTtsChunk = {
   /** Float32 PCM audio samples for this chunk, normalized in `[-1, 1]`. */
@@ -123,13 +125,43 @@ export type SupertonicTtsChunk = {
   readonly totalChunks: number;
 };
 
+/**
+ * Supertonic text-to-speech task runner.
+ * @category Speech / Types
+ * @typeParam K Voice style keys record constraint.
+ */
+export type SupertonicTextToSpeech<K extends PropertyKey = string> = {
+  /** Releases all allocated native models and static execution tensors. */
+  readonly dispose: () => void;
+
+  /**
+   * Streams synthesized audio chunks as an async generator as each text chunk
+   * finishes.
+   * @param text Input text string to synthesize into speech.
+   * @param options Per-call execution options.
+   * See {@link SupertonicTtsOptions}.
+   * @returns An AsyncGenerator yielding {@link SupertonicTtsChunk} audio
+   * buffers.
+   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if voice style is
+   * invalid or language is unsupported, `RESOURCE_BUSY` if the model is in use,
+   * or `RESOURCE_DISPOSED` if disposed.
+   */
+  readonly synthesize: (
+    text: string,
+    options: SupertonicTtsOptions<K>
+  ) => AsyncGenerator<SupertonicTtsChunk>;
+
+  /** Cancels any in-flight synthesis started by {@link synthesize}. */
+  readonly synthesizeStop: () => void;
+};
+
 /**
  * Creates a Supertonic Text-to-Speech pipeline.
  *
  * It validates all 4 sub-model method schemas, pre-allocates static execution
  * tensors, pre-parses voice styles into memory, and registers disposal hooks to
  * release all native resources.
- * @category Typescript API
+ * @category Speech / Tasks
  * @typeParam K Voice style keys record constraint.
  * @param config Supertonic TTS pipeline configuration containing model and asset paths.
  * See {@link SupertonicTtsModel}.
@@ -142,27 +174,7 @@ export type SupertonicTtsChunk = {
 export async function createSupertonicTextToSpeech<K extends PropertyKey>(
   config: SupertonicTtsModel<K>,
   runtime?: WorkletRuntime
-): Promise<{
-  /** Releases all allocated native models and static execution tensors. */
-  dispose: () => void;
-
-  /**
-   * Streams synthesized audio chunks as an async generator as each text chunk finishes.
-   * @param text Input text string to synthesize into speech.
-   * @param options Per-call execution options. See {@link SupertonicTtsOptions}.
-   * @returns An AsyncGenerator yielding {@link SupertonicTtsChunk} audio buffers.
-   * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if voice style is
-   * invalid or language is unsupported, `RESOURCE_BUSY` if the model is in use,
-   * or `RESOURCE_DISPOSED` if disposed.
-   */
-  synthesize: (
-    text: string,
-    options: SupertonicTtsOptions<K>
-  ) => AsyncGenerator<SupertonicTtsChunk>;
-
-  /** Cancels any in-flight synthesis started by {@link synthesize}. */
-  synthesizeStop: () => void;
-}> {
+): Promise<SupertonicTextToSpeech<K>> {
   const load = wrapAsync(loadModel, runtime);
   const [durationPredictor, vectorEstimator, textEncoder, vocoder] = await Promise.all([
     load(config.modelPaths.durationPredictor),
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
index 958112a7e0..4fd3f6458e 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
@@ -22,7 +22,7 @@ import { RnExecuTorchError } from '../../../core/error';
 
 /**
  * Sample rate (Hz) Whisper models expect their input waveform to be at.
- * @category Constants
+ * @category Speech / Constants
  */
 export const WHISPER_SAMPLE_RATE_HZ = 16000;
 
@@ -35,7 +35,7 @@ const BUFFER_SIZE = CHUNK_LENGTH_SECONDS * WHISPER_SAMPLE_RATE_HZ; // samples pe
 /**
  * Language codes supported by Whisper multilingual models. English-only
  * model variants only accept `'en'`.
- * @category Constants
+ * @category Speech / Constants
  */
 // prettier-ignore
 export const WHISPER_LANGUAGES = [
@@ -54,13 +54,13 @@ export const WHISPER_LANGUAGES = [
 /**
  * Union type of all language codes supported by Whisper. Derived from
  * {@link WHISPER_LANGUAGES}.
- * @category Types
+ * @category Speech / Types
  */
 export type WhisperLanguage = (typeof WHISPER_LANGUAGES)[number];
 
 /**
  * Options passed to a single transcription call.
- * @category Types
+ * @category Speech / Types
  */
 export type WhisperSttOptions<L extends WhisperLanguage = WhisperLanguage> = {
   /**
@@ -74,7 +74,7 @@ export type WhisperSttOptions<L extends WhisperLanguage = WhisperLanguage> = {
 /**
  * Options for the live-streaming transcription API.
  * Extends {@link WhisperSttOptions} with optional VAD tuning.
- * @category Types
+ * @category Speech / Types
  */
 export type WhisperStreamOptions<L extends WhisperLanguage = WhisperLanguage> =
   WhisperSttOptions<L> & {
@@ -87,7 +87,7 @@ export type WhisperStreamOptions<L extends WhisperLanguage = WhisperLanguage> =
 
 /**
  * Paths and metadata required to instantiate a Whisper speech-to-text model.
- * @category Types
+ * @category Speech / Types
  */
 export type WhisperSttModel<L extends WhisperLanguage = WhisperLanguage> = {
   /** Local path or remote URL of the `.pte` model. */
@@ -101,28 +101,15 @@ export type WhisperSttModel<L extends WhisperLanguage = WhisperLanguage> = {
 };
 
 /**
- * Loads a Whisper model and returns a set of transcription helpers.
- * @category Typescript API
- * @typeParam L Supported language codes constraint.
- * @param config Model paths, supported-language metadata, and VAD
- * configuration.
- * See {@link WhisperSttModel}.
- * @param runtime Optional worklet runtime thread on which to run the model
- * execution.
- * @returns A promise resolving to an object containing transcription and
- * disposal controls.
- * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model, tokenizer,
- * or VAD fails to load, or `SCHEMA_MISMATCH` if model schemas do not match the
- * Whisper specification.
+ * Whisper speech-to-text task runner.
+ * @category Speech / Types
+ * @typeParam L The language type accepted by the model.
  */
-export async function createWhisperSpeechToText<L extends WhisperLanguage = WhisperLanguage>(
-  config: WhisperSttModel<L>,
-  runtime?: WorkletRuntime
-): Promise<{
+export type WhisperSpeechToText<L extends WhisperLanguage = WhisperLanguage> = {
   /**
    * Releases all allocated native resources.
    */
-  dispose: () => void;
+  readonly dispose: () => void;
 
   /**
    * Asynchronously transcribes a pre-recorded mono waveform sampled at
@@ -136,7 +123,7 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
    * unsupported, `RESOURCE_BUSY` if the model is in use, or `RESOURCE_DISPOSED`
    * if disposed.
    */
-  transcribe: (
+  readonly transcribe: (
     audio: Float32Array,
     options: WhisperSttOptions<L>,
     onToken?: (token: string) => void
@@ -146,7 +133,7 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
    * Synchronous version of {@link transcribe} to be executed directly on the
    * caller or worklet thread.
    */
-  transcribeWorklet: (
+  readonly transcribeWorklet: (
     audio: Float32Array,
     options: WhisperSttOptions<L>,
     onToken?: (token: string) => void
@@ -155,7 +142,7 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
   /**
    * Interrupts and stops any active transcription call.
    */
-  transcribeStop: () => void;
+  readonly transcribeStop: () => void;
 
   /**
    * Async generator for real-time microphone transcription. Feed audio with
@@ -170,7 +157,7 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
    * unsupported, `RESOURCE_BUSY` if the model is in use, or `RESOURCE_DISPOSED`
    * if disposed.
    */
-  stream: (
+  readonly stream: (
     options: WhisperStreamOptions<L>
   ) => AsyncGenerator<{ committed: string; nonCommitted: string }>;
 
@@ -178,7 +165,7 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
    * Signals the {@link stream} generator to finalize the current segment and
    * return. Safe to call even when streaming is not active.
    */
-  streamStop: () => void;
+  readonly streamStop: () => void;
 
   /**
    * Appends a new PCM chunk to the live streaming buffer consumed by
@@ -186,8 +173,28 @@ export async function createWhisperSpeechToText<L extends WhisperLanguage = Whis
    * @param audioChunk The newly captured audio samples (16 kHz mono Float32
    * PCM).
    */
-  streamInsert: (audioChunk: Float32Array) => void;
-}> {
+  readonly streamInsert: (audioChunk: Float32Array) => void;
+};
+
+/**
+ * Loads a Whisper model and returns a set of transcription helpers.
+ * @category Speech / Tasks
+ * @typeParam L Supported language codes constraint.
+ * @param config Model paths, supported-language metadata, and VAD
+ * configuration.
+ * See {@link WhisperSttModel}.
+ * @param runtime Optional worklet runtime thread on which to run the model
+ * execution.
+ * @returns A promise resolving to an object containing transcription and
+ * disposal controls.
+ * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model, tokenizer,
+ * or VAD fails to load, or `SCHEMA_MISMATCH` if model schemas do not match the
+ * Whisper specification.
+ */
+export async function createWhisperSpeechToText<L extends WhisperLanguage = WhisperLanguage>(
+  config: WhisperSttModel<L>,
+  runtime?: WorkletRuntime
+): Promise<WhisperSpeechToText<L>> {
   const { modelPath, tokenizerPath, supportedLanguages, vadModel } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
   const tokenizer = await wrapAsync(loadTokenizer, runtime)(tokenizerPath);
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
index 42d964d7c2..cd60868603 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
@@ -5,13 +5,13 @@
 
 /**
  * Number of audio samples generated per single predicted duration tick.
- * @category Constants
+ * @category Speech / Constants
  */
 export const KOKORO_TICKS_PER_DURATION = 600;
 
 /**
  * Length of a single Kokoro voice reference vector (one row of a voice file).
- * @category Constants
+ * @category Speech / Constants
  */
 export const KOKORO_VOICE_REF_SIZE = 256;
 
@@ -39,7 +39,7 @@ const VOCAB: Record<string, number> = {
  * Silence (in milliseconds) appended after a chunk ending with a given phoneme,
  * so pauses between subsentences sound natural. Phonemes absent from the map
  * get no pause.
- * @category Constants
+ * @category Speech / Constants
  */
 // prettier-ignore
 export const KOKORO_PAUSE_MS: Record<string, number> = {
@@ -60,7 +60,7 @@ const BASE64_LOOKUP = /* @__PURE__ */ (() => {
  * Parses a base64-encoded Kokoro voice file into its raw float rows. Each row
  * holds a {@link KOKORO_VOICE_REF_SIZE}-long reference vector for one input
  * token count.
- * @category Utils
+ * @category Speech / Functions
  * @param base64 The base64 contents of the voice `.bin` file.
  * @returns The flattened voice matrix, row-major.
  */
@@ -91,7 +91,7 @@ export function parseVoice(base64: string): Float32Array {
 
 /**
  * Maps phonemes to vocabulary tokens, padded with the pad token on both ends.
- * @category Utils
+ * @category Speech / Functions
  * @param phonemes The phoneme sequence, split into code points.
  * @param totalLength The exact token count to produce, including padding.
  * @returns Token ids ready to be written into an `int64` tensor.
@@ -113,7 +113,7 @@ export function tokenize(phonemes: string[], totalLength: number): BigInt64Array
 /**
  * Scales per-token durations in place so that they sum up exactly to
  * `targetDuration`, distributing the rounding error by largest remainder.
- * @category Utils
+ * @category Speech / Functions
  * @param durations The per-token durations to scale in place.
  * @param targetDuration The exact sum the scaled durations must add up to.
  */
@@ -172,7 +172,7 @@ function findAudioBound(
 
 /**
  * Strips leading and trailing silence using a sliding-window moving average.
- * @category Utils
+ * @category Speech / Functions
  * @param audio The audio samples to strip.
  * @param margin The number of silence samples to preserve at each edge.
  * @param steps The moving average window length.
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts b/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
index 8a3cc2f465..264b24c9f7 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
@@ -10,13 +10,13 @@ declare const phonemizerBrand: unique symbol;
 
 /**
  * Union of all supported language codes in the G2P pipeline.
- * @category Types
+ * @category Speech / Types
  */
 export type PhonemizerLanguage = 'en-us' | 'en-gb' | 'fr' | 'es' | 'it' | 'pt' | 'de' | 'pl' | 'hi';
 
 /**
  * Configuration options and asset paths for initializing a {@link Phonemizer}.
- * @category Types
+ * @category Speech / Types
  */
 export type PhonemizerConfig = {
   /** Target language code to configure the G2P rules for. */
@@ -31,7 +31,7 @@ export type PhonemizerConfig = {
 
 /**
  * Native Grapheme-to-Phoneme (G2P) conversion interface.
- * @category Types
+ * @category Speech / Types
  */
 export type Phonemizer = {
   /**
@@ -58,7 +58,7 @@ export type Phonemizer = {
 
 /**
  * Creates a grapheme-to-phoneme (G2P) pipeline for the configured language.
- * @category Typescript API
+ * @category Speech / Functions
  * @param config Phonemizer configuration and asset paths.
  * See {@link PhonemizerConfig}.
  * @returns The native {@link Phonemizer} instance.
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
index fb99a2aa8a..e012267dc9 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
@@ -32,7 +32,7 @@ import { RnExecuTorchError } from '../../../core/error';
 
 /**
  * List of ISO language codes supported by Supertonic 3 for text synthesis conditioning.
- * @category Constants
+ * @category Speech / Constants
  */
 // prettier-ignore
 export const SUPERTONIC_SUPPORTED_LANGUAGES = [
@@ -43,7 +43,7 @@ export const SUPERTONIC_SUPPORTED_LANGUAGES = [
 
 /**
  * Supported Supertonic 3 ISO language code.
- * @category Types
+ * @category Speech / Types
  */
 export type SupertonicLanguage = (typeof SUPERTONIC_SUPPORTED_LANGUAGES)[number];
 
@@ -93,9 +93,9 @@ const DUPLICATE_QUOTES_PATTERN = /([`'""])\1+/g;
 const ENDING_PUNCTUATION_PATTERN = /[.!?;:,'")\]}…。」』】〉》›»]$/;
 
 /**
- * Normalizes unicode, replaces symbols/abbreviations, strips emojis, and cleans whitespace.
- * Should be run on full input text prior to chunk partitioning.
- * @category Utils
+ * Normalizes unicode, replaces symbols/abbreviations, strips emojis, and cleans
+ * whitespace. Should be run on full input text prior to chunk partitioning.
+ * @category Speech / Functions
  * @param text The raw input text.
  * @returns The normalized clean text.
  */
@@ -116,9 +116,9 @@ export function cleanText(text: string): string {
 }
 
 /**
- * Formats a single text chunk by ensuring ending punctuation and wrapping with language tags.
- * Should be run on individual partitioned text chunks.
- * @category Utils
+ * Formats a single text chunk by ensuring ending punctuation and wrapping with
+ * language tags. Should be run on individual partitioned text chunks.
+ * @category Speech / Functions
  * @param chunk The partitioned text chunk.
  * @param lang The language code.
  * @returns The formatted chunk ready for model input.
@@ -146,7 +146,7 @@ export function formatChunk(chunk: string, lang?: string): string {
 
 /**
  * Convenience wrapper combining {@link cleanText} and {@link formatChunk}.
- * @category Utils
+ * @category Speech / Functions
  * @param text The raw input text.
  * @param lang The language code.
  * @returns The preprocessed text.
@@ -159,7 +159,7 @@ export function preprocessText(text: string, lang?: string): string {
 /**
  * Encodes preprocessed text to character unicode index ids based on
  * unicode_indexer.json.
- * @category Utils
+ * @category Speech / Functions
  * @param text The preprocessed text.
  * @param indexer The unicode indexer character mapping array.
  * @returns BigInt64Array of character IDs.
@@ -177,7 +177,7 @@ export function encodeText(text: string, indexer: readonly number[]): BigInt64Ar
 
 /**
  * Parsed voice style tensors required by Supertonic 3 for style conditioning.
- * @category Types
+ * @category Speech / Types
  */
 export type SupertonicVoiceStyle = {
   /** Text-to-latent style embedding tensor data of shape [1, 50, 256] (12,800 floats). */
@@ -188,7 +188,7 @@ export type SupertonicVoiceStyle = {
 
 /**
  * Parses raw JSON voice style object data into Float32Arrays.
- * @category Utils
+ * @category Speech / Functions
  * @param json The parsed JSON voice style object.
  * @returns Parsed SupertonicVoiceStyle containing styleTtl and styleDp Float32Arrays.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the voice style JSON
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts b/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
index 1f6844f99f..c6aacbfe0f 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
@@ -28,7 +28,7 @@ const DEFAULT_SEPARATOR_PENALTY: Record<Tag, number> = { eos: 5, pause: 80, whit
 
 /**
  * Configuration options for text partitioning.
- * @category Types
+ * @category Speech / Types
  */
 export type PartitionOptions = {
   /**
@@ -93,7 +93,7 @@ function sliceAtCuts(text: string, cutIndices: number[]): string[] {
 /**
  * Divides input text into logical segments under the maximum character limit
  * using a forward dynamic programming algorithm.
- * @category Utils
+ * @category Speech / Functions
  * @param text The input text string to partition.
  * @param limit The maximum character limit per partition (must be >= 10).
  * @param options Optional configuration for TTFA prioritization and custom penalties.
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
index e563af89f0..42e3470e36 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
@@ -8,7 +8,7 @@ import { type Tensor } from '../../../core/tensor';
 
 /**
  * Options controlling how {@link extractFrames} slices and filters the waveform.
- * @category Types
+ * @category Speech / Types
  */
 export type ExtractFramesOptions = {
   /** Number of audio frames to extract and write into the destination tensor. */
@@ -27,7 +27,7 @@ export type ExtractFramesOptions = {
  * `dst` is fully zeroed first, so rows beyond `numFrames` stay zero (padding).
  * The frame length is taken from `hann`'s length and the padded width from
  * `dst`'s last dimension.
- * @category Typescript API
+ * @category Speech / Functions
  * @param waveform Input audio samples tensor. Expected 1D shape `[length]` with
  * data type `float32`. Framing starts at the first sample.
  * @param hann Precomputed Hann window tensor. Expected 1D shape `[frameLength]`
diff --git a/packages/react-native-executorch/src/fetcher/fetcher.ts b/packages/react-native-executorch/src/fetcher/fetcher.ts
index 8bbc9f963a..99f77e3988 100644
--- a/packages/react-native-executorch/src/fetcher/fetcher.ts
+++ b/packages/react-native-executorch/src/fetcher/fetcher.ts
@@ -19,7 +19,7 @@ const RNE_DIRECTORY = IS_ANDROID
 
 /**
  * Options controlling a {@link download} call.
- * @category Types
+ * @category Fetcher / Types
  */
 export interface DownloadOptions {
   /** Called with overall progress in `[0, 1]` as bytes arrive. */
@@ -409,7 +409,7 @@ function substituteRemoteSources<T>(node: T, resolved: ReadonlyMap<string, strin
  * streaming request that resumes an interrupted download from where it stopped.
  * When a config references several files, overall progress is weighted by their
  * byte sizes so a large model isn't reported the same as a tiny tokenizer.
- * @category Utils
+ * @category Fetcher / Functions
  * @typeParam T The shape of the value being resolved.
  * @param source A URL, a local path, or any nested object/array holding them.
  * @param options Progress and cancellation options.
diff --git a/packages/react-native-executorch/src/fetcher/telemetry.ts b/packages/react-native-executorch/src/fetcher/telemetry.ts
index 0aa9fdb26a..032755d6a9 100644
--- a/packages/react-native-executorch/src/fetcher/telemetry.ts
+++ b/packages/react-native-executorch/src/fetcher/telemetry.ts
@@ -24,7 +24,7 @@ function warn(message: string, error: unknown): void {
  * Analytics are enabled by default; call `setTelemetryEnabled(false)` (e.g. once
  * at app startup) to opt out. This does not affect the Hugging Face download
  * counter, a standard model-download stat that always fires.
- * @category Utils
+ * @category Fetcher / Functions
  * @param enabled Whether to send anonymous download analytics.
  */
 export function setTelemetryEnabled(enabled: boolean): void {
@@ -42,7 +42,7 @@ function isSwmHuggingFaceRepo(url: URL): boolean {
  * download-stats convention. No-op for any other host.
  *
  * See https://huggingface.co/docs/hub/models-download-stats
- * @category Utils
+ * @category Fetcher / Functions
  * @param uri The URI of the file being downloaded.
  */
 export function triggerHuggingFaceDownloadCounter(uri: string): void {
@@ -98,7 +98,7 @@ function getModelNameFromUri(uri: string): string {
  * Sends an anonymous download event to the Software Mansion analytics endpoint,
  * unless the app has opted out via {@link setTelemetryEnabled}. Fire-and-forget;
  * never throws and never blocks the download.
- * @category Utils
+ * @category Fetcher / Functions
  * @param uri The URI of the downloaded resource.
  */
 export function triggerDownloadEvent(uri: string): void {
diff --git a/packages/react-native-executorch/src/hooks/useResourceDownload.ts b/packages/react-native-executorch/src/hooks/useResourceDownload.ts
index 397fbe042b..44b467e72c 100644
--- a/packages/react-native-executorch/src/hooks/useResourceDownload.ts
+++ b/packages/react-native-executorch/src/hooks/useResourceDownload.ts
@@ -5,7 +5,7 @@ import { isRnExecuTorchError } from '../core/error';
 /**
  * Options accepted by {@link useResourceDownload} and by every `use<Task>` hook
  * built on top of it.
- * @category Types
+ * @category Hooks
  */
 export type ResourceOptions = {
   /** If true, prevents checks and downloads, resetting the hook state. */
diff --git a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
index 9c8dc16871..2aa020df8f 100644
--- a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
@@ -3,21 +3,18 @@ import { useResourceDownload, type ResourceOptions } from './useResourceDownload
 import {
   createKokoroTextToSpeech,
   type KokoroTtsModel,
+  type KokoroTextToSpeech,
 } from '../extensions/speech/tasks/kokoroTextToSpeech';
 import {
   createSupertonicTextToSpeech,
   type SupertonicTtsModel,
+  type SupertonicTextToSpeech,
 } from '../extensions/speech/tasks/supertonicTextToSpeech';
 
-type KokoroTts<K extends PropertyKey> = Awaited<
-  // prettier-ignore
-  ReturnType<typeof createKokoroTextToSpeech<K>>
->;
-type SupertonicTts<K extends PropertyKey> = Awaited<
-  // prettier-ignore
-  ReturnType<typeof createSupertonicTextToSpeech<K>>
->;
-
+/**
+ * Structured state and execution controls returned by {@link useTextToSpeech}.
+ * @category Hooks
+ */
 export type TtsHookResult<
   C,
   P extends {
@@ -59,7 +56,7 @@ export type TtsHookResult<
 export function useTextToSpeech<K extends PropertyKey>(
   config: KokoroTtsModel<K>,
   options?: ResourceOptions
-): TtsHookResult<KokoroTtsModel<K>, KokoroTts<K>>;
+): TtsHookResult<KokoroTtsModel<K>, KokoroTextToSpeech<K>>;
 
 /**
  * React hook to load and run the Supertonic 3 Text-to-Speech pipeline.
@@ -82,7 +79,7 @@ export function useTextToSpeech<K extends PropertyKey>(
 export function useTextToSpeech<K extends PropertyKey>(
   config: SupertonicTtsModel<K>,
   options?: ResourceOptions
-): TtsHookResult<SupertonicTtsModel<K>, SupertonicTts<K>>;
+): TtsHookResult<SupertonicTtsModel<K>, SupertonicTextToSpeech<K>>;
 
 export function useTextToSpeech<K extends PropertyKey>(
   config: KokoroTtsModel<K> | SupertonicTtsModel<K>,
@@ -92,7 +89,7 @@ export function useTextToSpeech<K extends PropertyKey>(
   // from that tag alone.
   const factory = (
     ttsConfig: KokoroTtsModel<K> | SupertonicTtsModel<K>
-  ): Promise<KokoroTts<K> | SupertonicTts<K>> => {
+  ): Promise<KokoroTextToSpeech<K> | SupertonicTextToSpeech<K>> => {
     switch (ttsConfig.name) {
       case 'kokoro':
         return createKokoroTextToSpeech(ttsConfig);
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 6d57dc7fab..930612d49b 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -71,7 +71,7 @@ export * from './fetcher';
 
 // Constants
 export { models } from './models';
-export * as constants from './constants';
+export * from './constants';
 
 // Task APIs — for developers needing manual lifetime/disposal control
 export * from './extensions/cv/tasks/classification';
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 105276527a..0f9cb693ad 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -1333,7 +1333,7 @@ const QWEN3_4B_XNNPACK_BF16: LLMModel = {
  * This provides Hugging Face repository URLs and baseline configurations for
  * tasks, allowing quick model loading and execution without manual option
  * setup.
- * @category Utils
+ * @category Models
  */
 export const models = {
   /**
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index 12413b981a..3298a5a15d 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -12,7 +12,7 @@ import RNBlobUtil from 'react-native-blob-util';
 /**
  * Retrieves the names of all ExecuTorch backends compiled and registered in the
  * native binary.
- * @category Utils
+ * @category Core / Functions
  * @returns An array of registered backend name strings (e.g. 'XnnpackBackend',
  * 'CoreMLBackend').
  */
@@ -21,6 +21,19 @@ export function getRegisteredBackends(): string[] {
   return rnexecutorchJsi.getExecuTorchRegisteredBackends();
 }
 
+/**
+ * Result of inspecting an ExecuTorch model file.
+ * @category Core / Types
+ */
+export type ModelInspection = {
+  /** The model source URL or file path that was inspected. */
+  readonly source: string;
+  /** Method signatures and tensor schema metadata. */
+  readonly schema: ModelSpec<ConcreteDim>;
+  /** Map of method names to the backends each method was compiled for. */
+  readonly backends: Record<string, readonly string[]>;
+};
+
 /**
  * Inspects an ExecuTorch model file to fetch its metadata and signature info
  * for all methods.
@@ -35,16 +48,12 @@ export function getRegisteredBackends(): string[] {
  * is not reused. Inspecting a remote model therefore re-downloads it on every
  * call and leaves nothing behind — call {@link download} first and inspect the
  * returned local path if you also intend to run the model.
- * @category Utils
+ * @category Core / Functions
  * @param source The remote HTTP URL or local path to the `.pte` model file.
  * @returns A promise resolving to an object containing the model source, method
  * signature metadata, and per-method backend usage.
  */
-export async function inspectModel(source: string): Promise<{
-  source: string;
-  schema: ModelSpec<ConcreteDim>;
-  backends: Record<string, readonly string[]>;
-}> {
+export async function inspectModel(source: string): Promise<ModelInspection> {
   let localPath = source;
   let downloaded = false;
 

From ed5d7b302bf06e86e0ed10b416f2a722089c01fb Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 14:32:20 +0200
Subject: [PATCH 20/38] refactor: improve JSDoc documentation, task types, and
 module exports

---
 .eslintrc.js                                  |   5 +-
 apps/computer-vision/app/detection/index.tsx  |   4 +-
 .../app/imageEmbeddings/index.tsx             |   2 +-
 apps/computer-vision/app/keypoint/index.tsx   |   2 +-
 apps/nlp/app/privacy-filter/index.tsx         |   8 +-
 apps/nlp/app/text-embeddings/index.tsx        |  22 ++-
 apps/speech/app/vad/index.tsx                 |   2 +-
 docs/typedoc.json                             |  30 +++-
 .../react-native-executorch/src/constants.ts  |   1 -
 .../react-native-executorch/src/core/error.ts |   1 -
 .../react-native-executorch/src/core/model.ts |   1 -
 .../src/core/runtime.ts                       |   1 -
 .../src/core/schema.ts                        |   1 -
 .../src/core/tensor.ts                        |   1 -
 .../src/extensions/cv/image.ts                |   1 -
 .../src/extensions/cv/index.ts                |   2 +-
 .../src/extensions/cv/ops/boxes.ts            |   1 -
 .../src/extensions/cv/ops/image.ts            |   1 -
 .../src/extensions/cv/ops/index.ts            |   1 -
 .../src/extensions/cv/ops/points.ts           |   1 -
 .../src/extensions/cv/ops/quad.ts             |   1 -
 .../src/extensions/cv/tasks/classification.ts |   1 -
 .../src/extensions/cv/tasks/imageEmbedding.ts |   1 -
 .../cv/tasks/instanceSegmentation.ts          |   1 -
 .../extensions/cv/tasks/keypointDetection.ts  |   1 -
 .../extensions/cv/tasks/objectDetection.ts    |   1 -
 .../src/extensions/cv/tasks/paddleOcr.ts      |   1 -
 .../src/extensions/cv/tasks/preprocessing.ts  |  10 +-
 .../extensions/cv/tasks/sdxsTextToImage.ts    |   5 +-
 .../cv/tasks/semanticSegmentation.ts          |   1 -
 .../src/extensions/cv/tasks/styleTransfer.ts  |   1 -
 .../src/extensions/cv/utils/paddleOcrUtils.ts |   1 -
 .../src/extensions/llm/index.ts               |   1 -
 .../src/extensions/llm/llmRunner.ts           |   3 +-
 .../extensions/llm/utils/chatPreprocessor.ts  |   1 -
 .../extensions/llm/utils/tokenizerConfig.ts   |   1 -
 .../src/extensions/llm/utils/toolCalling.ts   |   1 -
 .../src/extensions/math.ts                    |   1 -
 .../src/extensions/nlp/index.ts               |   1 -
 .../src/extensions/nlp/tasks/textEmbedding.ts |   1 -
 .../src/extensions/nlp/tasks/tokenization.ts  |  25 +--
 .../src/extensions/nlp/tokenizer.ts           |   1 -
 .../nlp/utils/privacyFilterUtils.ts           |   1 -
 .../src/extensions/speech/index.ts            |   1 -
 .../tasks/fsmnVoiceActivityDetection.ts       |  17 +-
 .../speech/tasks/kokoroTextToSpeech.ts        |   5 +-
 .../extensions/speech/utils/kokoroUtils.ts    |   1 -
 .../src/extensions/speech/utils/phonemizer.ts |   1 -
 .../speech/utils/supertonicUtils.ts           |   1 -
 .../speech/utils/textPartitioner.ts           |   1 -
 .../src/extensions/speech/utils/vadUtils.ts   |   1 -
 .../src/fetcher/fetcher.ts                    |   4 +-
 .../src/fetcher/telemetry.ts                  |   6 +-
 .../src/hooks/useResourceDownload.ts          |  18 +-
 .../src/hooks/useTextToSpeech.ts              |   6 +-
 packages/react-native-executorch/src/index.ts |  29 +++
 .../react-native-executorch/src/models.ts     | 168 +++++++++---------
 packages/react-native-executorch/src/utils.ts |  23 ++-
 58 files changed, 220 insertions(+), 212 deletions(-)

diff --git a/.eslintrc.js b/.eslintrc.js
index e6c4143afb..f2e119957a 100644
--- a/.eslintrc.js
+++ b/.eslintrc.js
@@ -10,8 +10,6 @@ const VALID_CATEGORIES = [
   'Core / Schema / Functions',
   'Core / Schema / Types',
   'Core / Types',
-  'Fetcher / Functions',
-  'Fetcher / Types',
   'Hooks',
   'LLM / Functions',
   'LLM / Tasks',
@@ -19,6 +17,7 @@ const VALID_CATEGORIES = [
   'Math / Functions',
   'Math / Types',
   'Models',
+  'Modules',
   'NLP / Constants',
   'NLP / Functions',
   'NLP / Tasks',
@@ -27,6 +26,8 @@ const VALID_CATEGORIES = [
   'Speech / Functions',
   'Speech / Tasks',
   'Speech / Types',
+  'Utils / Functions',
+  'Utils / Types',
 ];
 
 const CATEGORY_TAG_MATCH = `^(${VALID_CATEGORIES.join('|')})$`;
diff --git a/apps/computer-vision/app/detection/index.tsx b/apps/computer-vision/app/detection/index.tsx
index b2c990ea1c..fdee01d0a1 100644
--- a/apps/computer-vision/app/detection/index.tsx
+++ b/apps/computer-vision/app/detection/index.tsx
@@ -16,11 +16,11 @@ import { BoundingBox } from '../../components/BoundingBox';
 const MODEL_OPTIONS: ModelOption[] = [
   {
     label: 'SSDLite 320 MobileNet V3 Large (XNNPACK FP32)',
-    value: models.objectDetection.SSDLITE320_MOBILENET_V3_LARGE,
+    value: models.objectDetection.SSDLITE320_MOBILENET_V3_LARGE.DEFAULT,
   },
   {
     label: 'RF-DETR Nano (XNNPACK FP32)',
-    value: models.objectDetection.RFDETR_NANO,
+    value: models.objectDetection.RFDETR_NANO.DEFAULT,
   },
   {
     label: 'RF-DETR Nano (CoreML FP16)',
diff --git a/apps/computer-vision/app/imageEmbeddings/index.tsx b/apps/computer-vision/app/imageEmbeddings/index.tsx
index 47b8e04e7f..2d6dcf6323 100644
--- a/apps/computer-vision/app/imageEmbeddings/index.tsx
+++ b/apps/computer-vision/app/imageEmbeddings/index.tsx
@@ -71,7 +71,7 @@ function ImageEmbeddingsContent() {
   // Zero-shot classification pairs a CLIP image encoder with the CLIP text
   // encoder and scores the image against each text label by embedding similarity.
   const imageModel = useImageEmbedder(selectedImageModel);
-  const textModel = useTextEmbedder(models.textEmbeddings.CLIP_VIT_BASE_PATCH32_TEXT);
+  const textModel = useTextEmbedder(models.textEmbeddings.CLIP_VIT_BASE_PATCH32_TEXT.DEFAULT);
 
   const ready = imageModel.isReady && textModel.isReady;
 
diff --git a/apps/computer-vision/app/keypoint/index.tsx b/apps/computer-vision/app/keypoint/index.tsx
index 85a5260ee2..d0649996da 100644
--- a/apps/computer-vision/app/keypoint/index.tsx
+++ b/apps/computer-vision/app/keypoint/index.tsx
@@ -16,7 +16,7 @@ import { BoundingBox } from '../../components/BoundingBox';
 const MODEL_OPTIONS: ModelOption[] = [
   {
     label: 'BlazeFace (XNNPACK FP32)',
-    value: models.keypointDetection.BLAZEFACE,
+    value: models.keypointDetection.BLAZEFACE.DEFAULT,
   },
   {
     label: 'YOLO26 Pose (XNNPACK FP32)',
diff --git a/apps/nlp/app/privacy-filter/index.tsx b/apps/nlp/app/privacy-filter/index.tsx
index 9e9227ad1c..37ea7e8189 100644
--- a/apps/nlp/app/privacy-filter/index.tsx
+++ b/apps/nlp/app/privacy-filter/index.tsx
@@ -33,14 +33,18 @@ Reach her at maria.lopez@example.com or +1 (415) 555-0142. Address: 84 Cedar Hil
 /* cspell:enable */
 
 const MODELS: { label: string; value: PrivacyFilterModel; sample: string; iosOnly?: boolean }[] = [
-  { label: 'OpenAI (8 types)', value: models.privacyFilter.OPENAI, sample: OPENAI_SAMPLE },
+  { label: 'OpenAI (8 types)', value: models.privacyFilter.OPENAI.DEFAULT, sample: OPENAI_SAMPLE },
   {
     label: 'OpenAI MLX',
     value: models.privacyFilter.OPENAI.MLX_INT4,
     sample: OPENAI_SAMPLE,
     iosOnly: true,
   },
-  { label: 'Nemotron (55 types)', value: models.privacyFilter.NEMOTRON, sample: NEMOTRON_SAMPLE },
+  {
+    label: 'Nemotron (55 types)',
+    value: models.privacyFilter.NEMOTRON.DEFAULT,
+    sample: NEMOTRON_SAMPLE,
+  },
   {
     label: 'Nemotron MLX',
     value: models.privacyFilter.NEMOTRON.MLX_INT8,
diff --git a/apps/nlp/app/text-embeddings/index.tsx b/apps/nlp/app/text-embeddings/index.tsx
index 820f2296cf..7c33edc760 100644
--- a/apps/nlp/app/text-embeddings/index.tsx
+++ b/apps/nlp/app/text-embeddings/index.tsx
@@ -24,16 +24,22 @@ const MODELS: {
   docPrompt?: string;
   iosOnly?: boolean;
 }[] = [
-  { label: 'MiniLM L6', value: models.textEmbeddings.ALL_MINILM_L6_V2 },
-  { label: 'MPNet Base', value: models.textEmbeddings.ALL_MPNET_BASE_V2 },
-  { label: 'MultiQA MiniLM', value: models.textEmbeddings.MULTI_QA_MINILM_L6_COS_V1 },
-  { label: 'MultiQA MPNet', value: models.textEmbeddings.MULTI_QA_MPNET_BASE_DOT_V1 },
-  { label: 'Paraphrase ML', value: models.textEmbeddings.PARAPHRASE_MULTILINGUAL_MINILM_L12_V2 },
-  { label: 'DistilUSE ML', value: models.textEmbeddings.DISTILUSE_BASE_MULTILINGUAL_CASED_V2 },
-  { label: 'CLIP Text', value: models.textEmbeddings.CLIP_VIT_BASE_PATCH32_TEXT },
+  { label: 'MiniLM L6', value: models.textEmbeddings.ALL_MINILM_L6_V2.DEFAULT },
+  { label: 'MPNet Base', value: models.textEmbeddings.ALL_MPNET_BASE_V2.DEFAULT },
+  { label: 'MultiQA MiniLM', value: models.textEmbeddings.MULTI_QA_MINILM_L6_COS_V1.DEFAULT },
+  { label: 'MultiQA MPNet', value: models.textEmbeddings.MULTI_QA_MPNET_BASE_DOT_V1.DEFAULT },
+  {
+    label: 'Paraphrase ML',
+    value: models.textEmbeddings.PARAPHRASE_MULTILINGUAL_MINILM_L12_V2.DEFAULT,
+  },
+  {
+    label: 'DistilUSE ML',
+    value: models.textEmbeddings.DISTILUSE_BASE_MULTILINGUAL_CASED_V2.DEFAULT,
+  },
+  { label: 'CLIP Text', value: models.textEmbeddings.CLIP_VIT_BASE_PATCH32_TEXT.DEFAULT },
   {
     label: 'LFM2.5',
-    value: models.textEmbeddings.LFM2_5_EMBEDDING_350M,
+    value: models.textEmbeddings.LFM2_5_EMBEDDING_350M.DEFAULT,
     docPrompt: 'document: ',
   },
   {
diff --git a/apps/speech/app/vad/index.tsx b/apps/speech/app/vad/index.tsx
index f8a7e77d09..da3a47d563 100644
--- a/apps/speech/app/vad/index.tsx
+++ b/apps/speech/app/vad/index.tsx
@@ -15,7 +15,7 @@ const isSimulator = DeviceInfo.isEmulatorSync();
 
 function VADContent() {
   const { isReady, downloadProgress, error, detectVoiceOnStream, resetStream } =
-    useVoiceActivityDetector(models.voiceActivityDetection.FSMN_VAD);
+    useVoiceActivityDetector(models.voiceActivityDetection.FSMN_VAD.DEFAULT);
 
   const [isStreaming, setIsStreaming] = useState(false);
   const [isSpeaking, setIsSpeaking] = useState(false);
diff --git a/docs/typedoc.json b/docs/typedoc.json
index f68607dbf1..30c55c65bf 100644
--- a/docs/typedoc.json
+++ b/docs/typedoc.json
@@ -4,5 +4,33 @@
   "compilerOptions": {
     "skipLibCheck": true
   },
-  "excludeExternals": true
+  "excludeExternals": true,
+  "categoryOrder": [
+    "Hooks",
+    "Models",
+    "Modules",
+    "Core / Functions",
+    "Core / Types",
+    "Core / Errors",
+    "Core / Schema / Functions",
+    "Core / Schema / Types",
+    "Utils / Functions",
+    "Utils / Types",
+    "CV / Tasks",
+    "CV / Functions",
+    "CV / Types",
+    "CV / Constants",
+    "NLP / Tasks",
+    "NLP / Functions",
+    "NLP / Types",
+    "NLP / Constants",
+    "Speech / Tasks",
+    "Speech / Functions",
+    "Speech / Types",
+    "Speech / Constants",
+    "LLM / Tasks",
+    "LLM / Functions",
+    "LLM / Types",
+    "*"
+  ]
 }
diff --git a/packages/react-native-executorch/src/constants.ts b/packages/react-native-executorch/src/constants.ts
index 523361a026..518beaffde 100644
--- a/packages/react-native-executorch/src/constants.ts
+++ b/packages/react-native-executorch/src/constants.ts
@@ -3,7 +3,6 @@
  *
  * This module exports constant arrays, dataset labels, and normalization
  * parameters used across task pipelines and model configurations.
- * @module Constants
  */
 
 /**
diff --git a/packages/react-native-executorch/src/core/error.ts b/packages/react-native-executorch/src/core/error.ts
index 2d97b85a71..a3d2a529b3 100644
--- a/packages/react-native-executorch/src/core/error.ts
+++ b/packages/react-native-executorch/src/core/error.ts
@@ -8,7 +8,6 @@
  *
  * Use {@link isRnExecuTorchError} to safely narrow caught errors across
  * asynchronous calls and worklet runtime boundaries.
- * @module Core/Error
  */
 
 /**
diff --git a/packages/react-native-executorch/src/core/model.ts b/packages/react-native-executorch/src/core/model.ts
index a82bce3e3a..bc63ff04fb 100644
--- a/packages/react-native-executorch/src/core/model.ts
+++ b/packages/react-native-executorch/src/core/model.ts
@@ -4,7 +4,6 @@
  * Provides direct access to compiled `.pte` models in native C++ memory. Loaded
  * models expose synchronous method execution (`execute`) with pre-allocated
  * output buffers and manual native memory cleanup (`dispose`).
- * @module Core/Model
  */
 
 import { rnexecutorchJsi } from '../native/bridge';
diff --git a/packages/react-native-executorch/src/core/runtime.ts b/packages/react-native-executorch/src/core/runtime.ts
index 4fde38e09b..86b1bb62e8 100644
--- a/packages/react-native-executorch/src/core/runtime.ts
+++ b/packages/react-native-executorch/src/core/runtime.ts
@@ -4,7 +4,6 @@
  * Provides utilities to dispatch synchronous, heavy native operations (model
  * compilation, tensor inference) onto dedicated background worklet threads,
  * preventing them from blocking the React Native JavaScript thread.
- * @module Core/Runtime
  */
 
 import {
diff --git a/packages/react-native-executorch/src/core/schema.ts b/packages/react-native-executorch/src/core/schema.ts
index 229147a0ad..7ae0ba4aa4 100644
--- a/packages/react-native-executorch/src/core/schema.ts
+++ b/packages/react-native-executorch/src/core/schema.ts
@@ -66,7 +66,6 @@
  *    by passing `constant_methods={"get_model_schema": schema_json}` when
  *    lowering with `to_edge_transform_and_lower(...)` where `schema_json` is
  *    the JSON string encoding the model's `ModelSpec<ConcreteDim>`.
- * @module Core/Schema
  */
 import type { DType } from './tensor';
 import { RnExecuTorchError } from './error';
diff --git a/packages/react-native-executorch/src/core/tensor.ts b/packages/react-native-executorch/src/core/tensor.ts
index 07ef37027f..2fcbeb21ef 100644
--- a/packages/react-native-executorch/src/core/tensor.ts
+++ b/packages/react-native-executorch/src/core/tensor.ts
@@ -4,7 +4,6 @@
  * Tensors are the fundamental data structures used throughout React Native
  * ExecuTorch. They hold multidimensional typed arrays allocated in native
  * heap memory and provide transfers to and from JavaScript typed arrays.
- * @module Core/Tensor
  */
 
 import { rnexecutorchJsi } from '../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/cv/image.ts b/packages/react-native-executorch/src/extensions/cv/image.ts
index 3dc0d280e8..218974c7d9 100644
--- a/packages/react-native-executorch/src/extensions/cv/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/image.ts
@@ -1,6 +1,5 @@
 /**
  * Core image buffer types and pixel formats for Computer Vision.
- * @module CV/Image
  */
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/cv/index.ts b/packages/react-native-executorch/src/extensions/cv/index.ts
index ac585f06c3..a3d21e4a06 100644
--- a/packages/react-native-executorch/src/extensions/cv/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/index.ts
@@ -1,9 +1,9 @@
 /**
  * Computer Vision extension providing image buffers, spatial transformations,
  * OpenCV operators, bounding box utilities, and preprocessing pipelines.
- * @module CV
  */
 
 export * from './image';
 export * from './ops';
 export * from './utils/paddleOcrUtils';
+export * from './tasks/preprocessing';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index f18ab777e6..ac54e4c67c 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -1,7 +1,6 @@
 /**
  * Bounding box coordinate decoding, coordinate transforms, and Non-Maximum
  * Suppression (NMS).
- * @module CV/Ops/Boxes
  */
 
 import { rnexecutorchJsi } from '../../../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index d79ec64ff2..e15421db1b 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -4,7 +4,6 @@
  * Provides native OpenCV-accelerated image operations on tensors, including
  * spatial resizing, color space conversion, channel transposition (HWC/CHW),
  * pixel normalization, and colormap application.
- * @module CV/Ops/Image
  */
 
 import { rnexecutorchJsi } from '../../../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/index.ts b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
index ce242853b9..04a028b2fb 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
@@ -1,6 +1,5 @@
 /**
  * Aggregated image, box, and point operations.
- * @module CV Ops
  */
 export * as image from './image';
 export * as boxes from './boxes';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 05aac1b9c7..54b0b2587b 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -1,6 +1,5 @@
 /**
  * 2D point representation and spatial scaling utilities.
- * @module CV/Ops/Points
  */
 
 import type { ResizeMode } from './image';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
index 833fac8ef0..69c8c88487 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -1,7 +1,6 @@
 /**
  * Quadrilateral geometry, bounding box extraction, and image rectification
  * utilities.
- * @module CV/Ops/Quad
  */
 
 import { rnexecutorchJsi } from '../../../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
index b730cbff5b..f2e790ba17 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
@@ -1,7 +1,6 @@
 /**
  * Image classification task pipeline with integrated preprocessing and softmax
  * decoding.
- * @module CV/Tasks/Classification
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
index 7b0a3b7e2f..5e26ff6329 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
@@ -1,6 +1,5 @@
 /**
  * Image embedding and visual feature extraction task pipeline.
- * @module CV/Tasks/ImageEmbedding
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index 5831a30061..96379a19dc 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -1,7 +1,6 @@
 /**
  * Instance segmentation task pipeline with NMS, bounding box scaling, and mask
  * extraction.
- * @module CV/Tasks/InstanceSegmentation
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index 1cc5a340a9..54ebd50142 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -1,7 +1,6 @@
 /**
  * Keypoint and pose detection task pipeline with weighted NMS and landmark
  * scaling.
- * @module CV/Tasks/KeypointDetection
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index 0fc81e7132..70c0d7ae3f 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -1,7 +1,6 @@
 /**
  * Object detection task pipeline with integrated preprocessing, NMS, and box
  * scaling.
- * @module CV/Tasks/ObjectDetection
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 5b96c4b8fa..9c7ad3389b 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -1,7 +1,6 @@
 /**
  * PP-OCRv6 optical character recognition (DBNet text detection + SVTR text
  * recognition).
- * @module CV/Tasks/PaddleOcr
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
index b0b2cbb39e..203d676760 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
@@ -1,6 +1,5 @@
 /**
  * Reusable image preprocessing pipeline for neural network inputs.
- * @module CV/Tasks/Preprocessing
  */
 
 import { tensor, type Tensor } from '../../../core/tensor';
@@ -39,6 +38,11 @@ export type ImagePreprocessorOptions = {
  * @category CV / Types
  */
 export type ImagePreprocessor = {
+  /**
+   * Releases all allocated native resources.
+   */
+  readonly dispose: () => void;
+
   /**
    * Preprocesses the input image by resizing, converting color space, changing
    * format layout, and normalizing values, copying the output directly to the
@@ -51,10 +55,6 @@ export type ImagePreprocessor = {
    * data of shape `[3, H, W]` (or `[1, 3, H, W]`) and data type `float32`.
    */
   readonly process: (input: ImageBuffer) => Tensor;
-  /**
-   * Releases all allocated native resources.
-   */
-  readonly dispose: () => void;
 };
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
index 2e45714ee6..8b06f5ffbc 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
@@ -1,7 +1,6 @@
 /**
  * SDXS single-step text-to-image generation pipeline with latent diffusion and
  * TAESD decoding.
- * @module CV/Tasks/SdxsTextToImage
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
@@ -55,7 +54,9 @@ export type SdxsTextToImageModel = {
  * @category CV / Types
  */
 export type SdxsTextToImage = {
-  /** Releases all allocated native resources. */
+  /**
+   * Releases all allocated native resources.
+   */
   readonly dispose: () => void;
 
   /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
index 928518df82..2d4b8d3867 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
@@ -1,7 +1,6 @@
 /**
  * Semantic segmentation task pipeline with colormap mapping and output mask
  * rendering.
- * @module CV/Tasks/SemanticSegmentation
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
index 9a63fe031e..c14a8e0ef3 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
@@ -1,7 +1,6 @@
 /**
  * Neural style transfer task pipeline with output rendering and colorspace
  * conversion.
- * @module CV/Tasks/StyleTransfer
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
index af53fec8fd..388cf065dd 100644
--- a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
@@ -1,6 +1,5 @@
 /**
  * DBNet contour tracing and text quad extraction utilities for PP-OCRv6.
- * @module CV/Utils/PaddleOcrUtils
  */
 
 import { rnexecutorchJsi } from '../../../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/llm/index.ts b/packages/react-native-executorch/src/extensions/llm/index.ts
index fe61f5f582..bae6c85305 100644
--- a/packages/react-native-executorch/src/extensions/llm/index.ts
+++ b/packages/react-native-executorch/src/extensions/llm/index.ts
@@ -1,7 +1,6 @@
 /**
  * Large Language Model (LLM) runners, multi-turn chat sessions, and
  * tool-calling utilities.
- * @module LLM
  */
 
 export * from './llmRunner';
diff --git a/packages/react-native-executorch/src/extensions/llm/llmRunner.ts b/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
index 7681c31d3f..ede9ca1c85 100644
--- a/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
+++ b/packages/react-native-executorch/src/extensions/llm/llmRunner.ts
@@ -1,6 +1,5 @@
 /**
  * Low-level native ExecuTorch LLM runner types and factory.
- * @module LLM/LLMRunner
  */
 
 import type { Tensor } from '../../core/tensor';
@@ -103,7 +102,7 @@ export type LLMRunner = {
   readonly modalities: readonly Modality[];
 
   /**
-   * Disposes the native LLM runner and releases the loaded model memory.
+   * Releases all allocated native resources.
    */
   dispose(): void;
 
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
index e98545c753..8389657f8e 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
@@ -1,7 +1,6 @@
 /**
  * Multimodal chat message template rendering, incremental prompt diffing, and
  * media tensor preprocessing for LLM runners.
- * @module LLM/Utils/ChatPreprocessor
  */
 
 import { Template } from '@huggingface/jinja';
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
index 4b80a51b35..5f3f11d346 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/tokenizerConfig.ts
@@ -1,7 +1,6 @@
 /**
  * Parser for HuggingFace `tokenizer_config.json` chat templates and special
  * tokens.
- * @module LLM/Utils/TokenizerConfig
  */
 
 import { RnExecuTorchError } from '../../../core/error';
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
index 1ceec65d07..cbd968309f 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/toolCalling.ts
@@ -1,7 +1,6 @@
 /**
  * Tool call definition, XML/JSON parsing, and schema generation for LLM
  * function calling.
- * @module LLM/Utils/ToolCalling
  */
 
 import type { ChatMessageContent } from './chatPreprocessor';
diff --git a/packages/react-native-executorch/src/extensions/math.ts b/packages/react-native-executorch/src/extensions/math.ts
index fe10f7bac4..a3feccc823 100644
--- a/packages/react-native-executorch/src/extensions/math.ts
+++ b/packages/react-native-executorch/src/extensions/math.ts
@@ -1,6 +1,5 @@
 /**
  * Native C++ and JavaScript tensor math and random number generation utilities.
- * @module Math
  */
 
 import { rnexecutorchJsi } from '../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/nlp/index.ts b/packages/react-native-executorch/src/extensions/nlp/index.ts
index e25a66523f..476fe02d37 100644
--- a/packages/react-native-executorch/src/extensions/nlp/index.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/index.ts
@@ -1,7 +1,6 @@
 /**
  * Natural Language Processing (NLP) tokenization, text embedding, and privacy
  * filtering pipelines.
- * @module NLP
  */
 
 export * from './tokenizer';
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
index 8d0e4ca08c..bb7f5124a4 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
@@ -1,6 +1,5 @@
 /**
  * Text Embedding task pipeline for sentence transformers.
- * @module NLP/Tasks/TextEmbedding
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
index 9b0942ba82..bb506fe232 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
@@ -1,6 +1,5 @@
 /**
  * Tokenization task pipeline wrapping native HuggingFace tokenizers.
- * @module NLP/Tasks/Tokenization
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
@@ -8,25 +7,6 @@ import type { WorkletRuntime } from 'react-native-worklets';
 import { wrapAsync } from '../../../core/runtime';
 import { loadTokenizer } from '../tokenizer';
 
-/**
- * Asynchronous tokenizer task runner wrapping native HuggingFace tokenizer operations.
- * @category NLP / Types
- */
-export type Tokenizer = {
-  /** Encodes a string into token ids asynchronously. */
-  readonly encode: (text: string) => Promise<Int32Array>;
-  /** Decodes token ids back into a string asynchronously. */
-  readonly decode: (tokens: Int32Array, skipSpecialTokens?: boolean) => Promise<string>;
-  /** Returns the total vocabulary size. */
-  readonly getVocabSize: () => number;
-  /** Converts a numeric token id to its string piece representation. */
-  readonly idToToken: (id: number) => string;
-  /** Converts a string piece token to its numeric token id. */
-  readonly tokenToId: (token: string) => number;
-  /** Releases the native tokenizer resources. */
-  readonly dispose: () => void;
-};
-
 /**
  * Loads a tokenizer and exposes its operations with lifetime management for the
  * `useTokenizer` hook.
@@ -36,10 +16,7 @@ export type Tokenizer = {
  * @returns A promise resolving to the tokenizer operations and a `dispose`
  * handle that releases the native tokenizer.
  */
-export async function createTokenizer(
-  tokenizerPath: string,
-  runtime?: WorkletRuntime
-): Promise<Tokenizer> {
+export async function createTokenizer(tokenizerPath: string, runtime?: WorkletRuntime) {
   const tokenizer = await wrapAsync(loadTokenizer, runtime)(tokenizerPath);
   const dispose = () => tokenizer.dispose();
 
diff --git a/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts b/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
index 2690a9f0f2..3c7057b12c 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tokenizer.ts
@@ -1,6 +1,5 @@
 /**
  * Native HuggingFace fast tokenizer bindings.
- * @module NLP/Tokenizer
  */
 
 import { rnexecutorchJsi } from '../../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
index 5a20ce2885..e4e20601d6 100644
--- a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
@@ -1,6 +1,5 @@
 /**
  * BIOES grammar construction, Viterbi decoding, and PII entity span extraction.
- * @module NLP/Utils/PrivacyFilter
  */
 
 // Privacy-filter-specific helpers: BIOES grammar construction and constrained
diff --git a/packages/react-native-executorch/src/extensions/speech/index.ts b/packages/react-native-executorch/src/extensions/speech/index.ts
index eacd0f0174..2ada7825ea 100644
--- a/packages/react-native-executorch/src/extensions/speech/index.ts
+++ b/packages/react-native-executorch/src/extensions/speech/index.ts
@@ -1,7 +1,6 @@
 /**
  * Speech recognition, text-to-speech synthesis, phonemization, and audio
  * processing pipelines.
- * @module Speech
  */
 
 export * from './utils/vadUtils';
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
index 1e0f57812f..4c923e2981 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
@@ -1,6 +1,5 @@
 /**
  * Voice Activity Detection (VAD) task pipeline using the FSMN-VAD model.
- * @module Speech/Tasks/FsmnVoiceActivityDetection
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
@@ -36,7 +35,7 @@ const DEFAULT_DETECTION_MARGIN_MS = 100;
 
 /**
  * Tunable thresholds controlling how per-frame speech probabilities are turned
- * into speech {@link Segment}s.
+ * into speech {@link VadSegment}s.
  * @category Speech / Types
  */
 export type VadOptions = {
@@ -74,7 +73,7 @@ export type FsmnVadModel = {
  * A detected speech region, with start and end expressed in seconds.
  * @category Speech / Types
  */
-export type Segment = {
+export type VadSegment = {
   /** Start time of the speech segment in seconds. */
   readonly start: number;
   /** End time of the speech segment in seconds. */
@@ -115,7 +114,7 @@ function hannWindow(size: number): Float32Array {
 // threshold with hysteresis, pad both ends, then merge near-adjacent regions.
 // `scores[i]` holds the non-speech probability of frame `i`, so the speech
 // probability is `1 - scores[i]`.
-function postprocess(scores: Float32Array, options: Required<VadOptions>): Segment[] {
+function postprocess(scores: Float32Array, options: Required<VadOptions>): VadSegment[] {
   'worklet';
   const threshold = options.speechThreshold;
   const minSpeechHops = Math.floor(options.minSpeechDurationMs / HOP_LENGTH_MS);
@@ -126,7 +125,7 @@ function postprocess(scores: Float32Array, options: Required<VadOptions>): Segme
   // Threshold with hysteresis: a region must stay above the threshold for
   // `minSpeechHops` to open a segment, and below it for `minSilenceHops` to
   // close one. Bounds are in hop (frame) units.
-  const hops: Segment[] = [];
+  const hops: VadSegment[] = [];
   let triggered = false;
   let startHop = -1;
   let potentialStart = -1;
@@ -157,7 +156,7 @@ function postprocess(scores: Float32Array, options: Required<VadOptions>): Segme
   if (triggered) hops.push({ start: startHop, end: scores.length });
 
   // Pad both ends, then merge regions separated by at most `mergeGapMs`.
-  const merged: Segment[] = [];
+  const merged: VadSegment[] = [];
   for (const hop of hops) {
     const start = hop.start > speechPadHops ? hop.start - speechPadHops : 0;
     const end = Math.min(hop.end + speechPadHops, scores.length);
@@ -193,13 +192,13 @@ export type FsmnVoiceActivityDetector = {
    * @throws {RnExecuTorchError} With code `RESOURCE_BUSY` if the model is in
    * use, or `RESOURCE_DISPOSED` if disposed.
    */
-  readonly detectVoice: (waveform: Float32Array, options?: VadOptions) => Promise<Segment[]>;
+  readonly detectVoice: (waveform: Float32Array, options?: VadOptions) => Promise<VadSegment[]>;
 
   /**
    * Synchronous version of {@link detectVoice} to be executed directly on the caller
    * or worklet thread.
    */
-  readonly detectVoiceWorklet: (waveform: Float32Array, options?: VadOptions) => Segment[];
+  readonly detectVoiceWorklet: (waveform: Float32Array, options?: VadOptions) => VadSegment[];
 
   /**
    * Appends a live audio chunk to a bounded rolling window, runs detection over
@@ -275,7 +274,7 @@ export async function createFsmnVoiceActivityDetector(
     model.dispose();
   };
 
-  const detectVoiceWorklet = (waveform: Float32Array, options?: VadOptions): Segment[] => {
+  const detectVoiceWorklet = (waveform: Float32Array, options?: VadOptions): VadSegment[] => {
     'worklet';
     const mergedOpts: Required<VadOptions> = { ...defaultOptions, ...options };
     const numFrames = Math.floor((waveform.length - FRAME_LENGTH) / HOP_LENGTH);
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
index 1b117c1d33..7cfbcaad8c 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
@@ -1,6 +1,5 @@
 /**
  * Kokoro Text-to-Speech (TTS) synthesis task pipeline.
- * @module Speech/Tasks/KokoroTextToSpeech
  */
 
 import type { WorkletRuntime } from 'react-native-worklets';
@@ -112,7 +111,9 @@ export type KokoroTtsChunk = {
  * @typeParam K Voice keys record constraint.
  */
 export type KokoroTextToSpeech<K extends PropertyKey = string> = {
-  /** Releases the allocated native models, phonemizer and execution tensors. */
+  /**
+   * Releases all allocated native resources.
+   */
   readonly dispose: () => void;
 
   /**
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
index cd60868603..67c1b164a2 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/kokoroUtils.ts
@@ -1,6 +1,5 @@
 /**
  * Utility functions, vocoder constants, and voice decoders for Kokoro TTS.
- * @module Speech/Utils/Kokoro
  */
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts b/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
index 264b24c9f7..e6d0b56d86 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/phonemizer.ts
@@ -1,6 +1,5 @@
 /**
  * Grapheme-to-phoneme (G2P) conversion using native Phonemis bindings.
- * @module Speech/Utils/Phonemizer
  */
 
 import { rnexecutorchJsi } from '../../../native/bridge';
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
index e012267dc9..5c758f2b87 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/supertonicUtils.ts
@@ -1,7 +1,6 @@
 /**
  * Text normalization, chunk formatting, and voice style parsing utilities for
  * Supertonic TTS.
- * @module Speech/Utils/Supertonic
  */
 
 import { RnExecuTorchError } from '../../../core/error';
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts b/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
index c6aacbfe0f..bf83f63bc3 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/textPartitioner.ts
@@ -1,7 +1,6 @@
 /**
  * Dynamic programming text partitioning algorithm for splitting long sentences
  * into natural speech chunks.
- * @module Speech/Utils/TextPartitioner
  */
 
 import { RnExecuTorchError } from '../../../core/error';
diff --git a/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts b/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
index 42e3470e36..39fdf917f5 100644
--- a/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
+++ b/packages/react-native-executorch/src/extensions/speech/utils/vadUtils.ts
@@ -1,6 +1,5 @@
 /**
  * Audio framing and pre-emphasis feature extraction for Voice Activity Detection.
- * @module Speech/Utils/VAD
  */
 
 import { rnexecutorchJsi } from '../../../native/bridge';
diff --git a/packages/react-native-executorch/src/fetcher/fetcher.ts b/packages/react-native-executorch/src/fetcher/fetcher.ts
index 99f77e3988..32f12adcf3 100644
--- a/packages/react-native-executorch/src/fetcher/fetcher.ts
+++ b/packages/react-native-executorch/src/fetcher/fetcher.ts
@@ -19,7 +19,7 @@ const RNE_DIRECTORY = IS_ANDROID
 
 /**
  * Options controlling a {@link download} call.
- * @category Fetcher / Types
+ * @category Utils / Types
  */
 export interface DownloadOptions {
   /** Called with overall progress in `[0, 1]` as bytes arrive. */
@@ -409,7 +409,7 @@ function substituteRemoteSources<T>(node: T, resolved: ReadonlyMap<string, strin
  * streaming request that resumes an interrupted download from where it stopped.
  * When a config references several files, overall progress is weighted by their
  * byte sizes so a large model isn't reported the same as a tiny tokenizer.
- * @category Fetcher / Functions
+ * @category Utils / Functions
  * @typeParam T The shape of the value being resolved.
  * @param source A URL, a local path, or any nested object/array holding them.
  * @param options Progress and cancellation options.
diff --git a/packages/react-native-executorch/src/fetcher/telemetry.ts b/packages/react-native-executorch/src/fetcher/telemetry.ts
index 032755d6a9..0a79a16760 100644
--- a/packages/react-native-executorch/src/fetcher/telemetry.ts
+++ b/packages/react-native-executorch/src/fetcher/telemetry.ts
@@ -24,7 +24,7 @@ function warn(message: string, error: unknown): void {
  * Analytics are enabled by default; call `setTelemetryEnabled(false)` (e.g. once
  * at app startup) to opt out. This does not affect the Hugging Face download
  * counter, a standard model-download stat that always fires.
- * @category Fetcher / Functions
+ * @category Utils / Functions
  * @param enabled Whether to send anonymous download analytics.
  */
 export function setTelemetryEnabled(enabled: boolean): void {
@@ -42,7 +42,7 @@ function isSwmHuggingFaceRepo(url: URL): boolean {
  * download-stats convention. No-op for any other host.
  *
  * See https://huggingface.co/docs/hub/models-download-stats
- * @category Fetcher / Functions
+ * @category Utils / Functions
  * @param uri The URI of the file being downloaded.
  */
 export function triggerHuggingFaceDownloadCounter(uri: string): void {
@@ -98,7 +98,7 @@ function getModelNameFromUri(uri: string): string {
  * Sends an anonymous download event to the Software Mansion analytics endpoint,
  * unless the app has opted out via {@link setTelemetryEnabled}. Fire-and-forget;
  * never throws and never blocks the download.
- * @category Fetcher / Functions
+ * @category Utils / Functions
  * @param uri The URI of the downloaded resource.
  */
 export function triggerDownloadEvent(uri: string): void {
diff --git a/packages/react-native-executorch/src/hooks/useResourceDownload.ts b/packages/react-native-executorch/src/hooks/useResourceDownload.ts
index 44b467e72c..33b584b0bc 100644
--- a/packages/react-native-executorch/src/hooks/useResourceDownload.ts
+++ b/packages/react-native-executorch/src/hooks/useResourceDownload.ts
@@ -2,21 +2,9 @@ import { useState, useEffect, useMemo } from 'react';
 import { download } from '../fetcher/fetcher';
 import { isRnExecuTorchError } from '../core/error';
 
-/**
- * Options accepted by {@link useResourceDownload} and by every `use<Task>` hook
- * built on top of it.
- * @category Hooks
- */
-export type ResourceOptions = {
-  /** If true, prevents checks and downloads, resetting the hook state. */
-  preventLoad?: boolean;
-  /**
-   * Re-downloads every remote source even when it is already cached, replacing
-   * the cached copy. Use to recover from a corrupted file or to pick up a model
-   * that changed behind a stable URL.
-   */
-  forceDownload?: boolean;
-};
+import type { ResourceOptions } from '../utils';
+
+export type { ResourceOptions };
 
 /**
  * React hook to manage downloading and local caching of the remote resources
diff --git a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
index 2aa020df8f..44b0f849e5 100644
--- a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
@@ -11,11 +11,7 @@ import {
   type SupertonicTextToSpeech,
 } from '../extensions/speech/tasks/supertonicTextToSpeech';
 
-/**
- * Structured state and execution controls returned by {@link useTextToSpeech}.
- * @category Hooks
- */
-export type TtsHookResult<
+type TtsHookResult<
   C,
   P extends {
     synthesize: (...args: any[]) => any;
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 930612d49b..2b8ba5b91b 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -83,6 +83,7 @@ export * from './extensions/cv/tasks/objectDetection';
 export * from './extensions/cv/tasks/imageEmbedding';
 export * from './extensions/cv/tasks/sdxsTextToImage';
 export * from './extensions/cv/tasks/paddleOcr';
+export * from './extensions/cv/tasks/preprocessing';
 export * from './extensions/llm/tasks/llmChatSession';
 export * from './extensions/nlp/tasks/tokenization';
 export * from './extensions/nlp/tasks/textEmbedding';
@@ -99,12 +100,40 @@ export * from './core/tensor';
 export * from './core/runtime';
 
 export type * from './core/schema';
+/**
+ * Model schema validation and dimension constraints.
+ * @category Modules
+ */
 export * as schema from './core/schema';
 
+/**
+ * Mathematical tensor operations and activation functions.
+ * @category Modules
+ */
 export * as math from './extensions/math';
+
+/**
+ * Computer vision operations, geometry helpers, and image preprocessing.
+ * @category Modules
+ */
 export * as cv from './extensions/cv';
+
+/**
+ * LLM runner, tokenizer configurations, and multimodal chat preprocessing.
+ * @category Modules
+ */
 export * as llm from './extensions/llm';
+
+/**
+ * NLP tokenizers and privacy filter utilities.
+ * @category Modules
+ */
 export * as nlp from './extensions/nlp';
+
+/**
+ * Speech utilities, phonemizers, text partitioning, and VAD framing.
+ * @category Modules
+ */
 export * as speech from './extensions/speech';
 
 // Utils
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 0f9cb693ad..2f0bfafea8 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -1348,7 +1348,7 @@ export const models = {
      * classification.
      */
     EFFICIENTNET_V2_S: {
-      ...EFFICIENTNET_V2_S_XNNPACK_INT8,
+      DEFAULT: EFFICIENTNET_V2_S_XNNPACK_INT8,
       XNNPACK_INT8: EFFICIENTNET_V2_S_XNNPACK_INT8,
       XNNPACK_FP32: EFFICIENTNET_V2_S_XNNPACK_FP32,
       COREML_FP16: EFFICIENTNET_V2_S_COREML_FP16,
@@ -1365,7 +1365,7 @@ export const models = {
      * style effect.
      */
     CANDY: {
-      ...STYLE_TRANSFER_CANDY_XNNPACK_INT8,
+      DEFAULT: STYLE_TRANSFER_CANDY_XNNPACK_INT8,
       XNNPACK_FP32: STYLE_TRANSFER_CANDY_XNNPACK_FP32,
       XNNPACK_INT8: STYLE_TRANSFER_CANDY_XNNPACK_INT8,
       COREML_FP16: STYLE_TRANSFER_CANDY_COREML_FP16,
@@ -1375,7 +1375,7 @@ export const models = {
      * pattern.
      */
     MOSAIC: {
-      ...STYLE_TRANSFER_MOSAIC_XNNPACK_INT8,
+      DEFAULT: STYLE_TRANSFER_MOSAIC_XNNPACK_INT8,
       XNNPACK_FP32: STYLE_TRANSFER_MOSAIC_XNNPACK_FP32,
       XNNPACK_INT8: STYLE_TRANSFER_MOSAIC_XNNPACK_INT8,
       COREML_FP16: STYLE_TRANSFER_MOSAIC_COREML_FP16,
@@ -1385,7 +1385,7 @@ export const models = {
      * painting aesthetic.
      */
     RAIN_PRINCESS: {
-      ...STYLE_TRANSFER_RAIN_PRINCESS_XNNPACK_INT8,
+      DEFAULT: STYLE_TRANSFER_RAIN_PRINCESS_XNNPACK_INT8,
       XNNPACK_FP32: STYLE_TRANSFER_RAIN_PRINCESS_XNNPACK_FP32,
       XNNPACK_INT8: STYLE_TRANSFER_RAIN_PRINCESS_XNNPACK_INT8,
       COREML_FP16: STYLE_TRANSFER_RAIN_PRINCESS_COREML_FP16,
@@ -1395,7 +1395,7 @@ export const models = {
      * abstract art style.
      */
     UDNIE: {
-      ...STYLE_TRANSFER_UDNIE_XNNPACK_INT8,
+      DEFAULT: STYLE_TRANSFER_UDNIE_XNNPACK_INT8,
       XNNPACK_FP32: STYLE_TRANSFER_UDNIE_XNNPACK_FP32,
       XNNPACK_INT8: STYLE_TRANSFER_UDNIE_XNNPACK_INT8,
       COREML_FP16: STYLE_TRANSFER_UDNIE_COREML_FP16,
@@ -1413,7 +1413,7 @@ export const models = {
      * Ideal for background blur and replacement effects.
      */
     SELFIE_SEGMENTATION: {
-      ...SELFIE_SEGMENTATION_XNNPACK_FP32,
+      DEFAULT: SELFIE_SEGMENTATION_XNNPACK_FP32,
       XNNPACK_FP32: SELFIE_SEGMENTATION_XNNPACK_FP32,
       COREML_FP16: SELFIE_SEGMENTATION_COREML_FP16,
     },
@@ -1422,7 +1422,7 @@ export const models = {
      * 256x144 checkpoint rather than a resize of the portrait model.
      */
     SELFIE_SEGMENTATION_LANDSCAPE: {
-      ...SELFIE_SEGMENTATION_LANDSCAPE_XNNPACK_FP32,
+      DEFAULT: SELFIE_SEGMENTATION_LANDSCAPE_XNNPACK_FP32,
       XNNPACK_FP32: SELFIE_SEGMENTATION_LANDSCAPE_XNNPACK_FP32,
       COREML_FP16: SELFIE_SEGMENTATION_LANDSCAPE_COREML_FP16,
     },
@@ -1432,7 +1432,7 @@ export const models = {
      * real-time pixel-level segmentation on mobile devices.
      */
     LRASPP_MOBILENET_V3_LARGE: {
-      ...LRASPP_MOBILENET_V3_LARGE_XNNPACK_INT8,
+      DEFAULT: LRASPP_MOBILENET_V3_LARGE_XNNPACK_INT8,
       XNNPACK_FP32: LRASPP_MOBILENET_V3_LARGE_XNNPACK_FP32,
       XNNPACK_INT8: LRASPP_MOBILENET_V3_LARGE_XNNPACK_INT8,
       COREML_FP16: LRASPP_MOBILENET_V3_LARGE_COREML_FP16,
@@ -1443,7 +1443,7 @@ export const models = {
      * utilizing atrous spatial pyramid pooling.
      */
     DEEPLAB_V3_RESNET50: {
-      ...DEEPLAB_V3_RESNET50_XNNPACK_INT8,
+      DEFAULT: DEEPLAB_V3_RESNET50_XNNPACK_INT8,
       XNNPACK_FP32: DEEPLAB_V3_RESNET50_XNNPACK_FP32,
       XNNPACK_INT8: DEEPLAB_V3_RESNET50_XNNPACK_INT8,
       COREML_FP16: DEEPLAB_V3_RESNET50_COREML_FP16,
@@ -1454,7 +1454,7 @@ export const models = {
      * maximum segmentation detail and boundary accuracy.
      */
     DEEPLAB_V3_RESNET101: {
-      ...DEEPLAB_V3_RESNET101_XNNPACK_INT8,
+      DEFAULT: DEEPLAB_V3_RESNET101_XNNPACK_INT8,
       XNNPACK_FP32: DEEPLAB_V3_RESNET101_XNNPACK_FP32,
       XNNPACK_INT8: DEEPLAB_V3_RESNET101_XNNPACK_INT8,
       COREML_FP16: DEEPLAB_V3_RESNET101_COREML_FP16,
@@ -1465,7 +1465,7 @@ export const models = {
      * extraction quality with a lightweight mobile backbone.
      */
     DEEPLAB_V3_MOBILENET_V3_LARGE: {
-      ...DEEPLAB_V3_MOBILENET_V3_LARGE_XNNPACK_INT8,
+      DEFAULT: DEEPLAB_V3_MOBILENET_V3_LARGE_XNNPACK_INT8,
       XNNPACK_FP32: DEEPLAB_V3_MOBILENET_V3_LARGE_XNNPACK_FP32,
       XNNPACK_INT8: DEEPLAB_V3_MOBILENET_V3_LARGE_XNNPACK_INT8,
       COREML_FP16: DEEPLAB_V3_MOBILENET_V3_LARGE_COREML_FP16,
@@ -1475,7 +1475,7 @@ export const models = {
      * ResNet-50 backbone (21 classes, see {@link PASCAL_VOC_LABELS}).
      */
     FCN_RESNET50: {
-      ...FCN_RESNET50_XNNPACK_INT8,
+      DEFAULT: FCN_RESNET50_XNNPACK_INT8,
       XNNPACK_FP32: FCN_RESNET50_XNNPACK_FP32,
       XNNPACK_INT8: FCN_RESNET50_XNNPACK_INT8,
       COREML_FP16: FCN_RESNET50_COREML_FP16,
@@ -1485,7 +1485,7 @@ export const models = {
      * ResNet-101 backbone (21 classes, see {@link PASCAL_VOC_LABELS}).
      */
     FCN_RESNET101: {
-      ...FCN_RESNET101_XNNPACK_INT8,
+      DEFAULT: FCN_RESNET101_XNNPACK_INT8,
       XNNPACK_FP32: FCN_RESNET101_XNNPACK_FP32,
       XNNPACK_INT8: FCN_RESNET101_XNNPACK_INT8,
       COREML_FP16: FCN_RESNET101_COREML_FP16,
@@ -1502,7 +1502,7 @@ export const models = {
      * detector suited for real-time mobile applications.
      */
     SSDLITE320_MOBILENET_V3_LARGE: {
-      ...SSDLITE320_MOBILENET_V3_LARGE_XNNPACK_FP32,
+      DEFAULT: SSDLITE320_MOBILENET_V3_LARGE_XNNPACK_FP32,
       XNNPACK_FP32: SSDLITE320_MOBILENET_V3_LARGE_XNNPACK_FP32,
       COREML_FP16: SSDLITE320_MOBILENET_V3_LARGE_COREML_FP16,
     },
@@ -1512,7 +1512,7 @@ export const models = {
      * object detector.
      */
     RFDETR_NANO: {
-      ...RFDETR_NANO_DETECTOR_XNNPACK_FP32,
+      DEFAULT: RFDETR_NANO_DETECTOR_XNNPACK_FP32,
       XNNPACK_FP32: RFDETR_NANO_DETECTOR_XNNPACK_FP32,
       COREML_FP16: RFDETR_NANO_DETECTOR_COREML_FP16,
     },
@@ -1523,13 +1523,13 @@ export const models = {
      * 512x512, 640x640).
      */
     YOLO26: {
-      ...YOLO26_NANO_384_XNNPACK_FP32,
+      DEFAULT: YOLO26_NANO_384_XNNPACK_FP32,
       /**
        * Nano scale YOLO26 object detection model. High speed, ultra low
        * latency.
        */
       NANO: {
-        ...YOLO26_NANO_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_NANO_384_XNNPACK_FP32,
         SIZE_384: {
           XNNPACK_FP32: YOLO26_NANO_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_NANO_384_COREML_FP16,
@@ -1548,7 +1548,7 @@ export const models = {
        * accuracy.
        */
       SMALL: {
-        ...YOLO26_SMALL_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_SMALL_384_XNNPACK_FP32,
         SIZE_384: {
           XNNPACK_FP32: YOLO26_SMALL_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_SMALL_384_COREML_FP16,
@@ -1567,7 +1567,7 @@ export const models = {
        * complex scenes.
        */
       MEDIUM: {
-        ...YOLO26_MEDIUM_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_MEDIUM_384_XNNPACK_FP32,
         SIZE_384: {
           XNNPACK_FP32: YOLO26_MEDIUM_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_MEDIUM_384_COREML_FP16,
@@ -1585,7 +1585,7 @@ export const models = {
        * Large scale YOLO26 object detection model. High accuracy model variant.
        */
       LARGE: {
-        ...YOLO26_LARGE_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_LARGE_384_XNNPACK_FP32,
         SIZE_384: {
           XNNPACK_FP32: YOLO26_LARGE_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_LARGE_384_COREML_FP16,
@@ -1604,7 +1604,7 @@ export const models = {
        * performance.
        */
       XLARGE: {
-        ...YOLO26_XLARGE_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_XLARGE_384_XNNPACK_FP32,
         SIZE_384: {
           XNNPACK_FP32: YOLO26_XLARGE_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_XLARGE_384_COREML_FP16,
@@ -1632,7 +1632,7 @@ export const models = {
      * {@link BLAZEFACE_LANDMARKS}).
      */
     BLAZEFACE: {
-      ...BLAZEFACE_XNNPACK_FP32,
+      DEFAULT: BLAZEFACE_XNNPACK_FP32,
       XNNPACK_FP32: BLAZEFACE_XNNPACK_FP32,
     },
     /**
@@ -1641,7 +1641,7 @@ export const models = {
      * resolutions.
      */
     YOLO26_POSE: {
-      ...YOLO26_POSE_384_XNNPACK_FP32,
+      DEFAULT: YOLO26_POSE_384_XNNPACK_FP32,
       SIZE_384: { XNNPACK_FP32: YOLO26_POSE_384_XNNPACK_FP32 },
       SIZE_512: { XNNPACK_FP32: YOLO26_POSE_512_XNNPACK_FP32 },
       SIZE_640: { XNNPACK_FP32: YOLO26_POSE_640_XNNPACK_FP32 },
@@ -1651,7 +1651,7 @@ export const models = {
      * predicting 17 COCO body keypoints (see {@link COCO_LANDMARKS}).
      */
     RFDETR_KEYPOINT: {
-      ...RFDETR_KEYPOINT_XNNPACK_FP32,
+      DEFAULT: RFDETR_KEYPOINT_XNNPACK_FP32,
       XNNPACK_FP32: RFDETR_KEYPOINT_XNNPACK_FP32,
       COREML_FP32: RFDETR_KEYPOINT_COREML_FP32,
       MLX_FP32: RFDETR_KEYPOINT_MLX_FP32,
@@ -1673,7 +1673,7 @@ export const models = {
        * FastSAM Small - lightweight instance segmenter for mobile.
        */
       S: {
-        ...FASTSAM_S_XNNPACK_FP32,
+        DEFAULT: FASTSAM_S_XNNPACK_FP32,
         XNNPACK_FP32: FASTSAM_S_XNNPACK_FP32,
         COREML_FP16: FASTSAM_S_COREML_FP16,
       },
@@ -1681,7 +1681,7 @@ export const models = {
        * FastSAM Extra Large - high-accuracy instance segmenter.
        */
       X: {
-        ...FASTSAM_X_XNNPACK_FP32,
+        DEFAULT: FASTSAM_X_XNNPACK_FP32,
         XNNPACK_FP32: FASTSAM_X_XNNPACK_FP32,
         COREML_FP16: FASTSAM_X_COREML_FP16,
       },
@@ -1692,7 +1692,7 @@ export const models = {
      * {@link COCO_CLASSES}).
      */
     RFDETR_NANO: {
-      ...RFDETR_NANO_SEG_COREML_FP16,
+      DEFAULT: RFDETR_NANO_SEG_COREML_FP16,
       COREML_FP16: RFDETR_NANO_SEG_COREML_FP16,
       XNNPACK_FP32: RFDETR_NANO_SEG_XNNPACK_FP32,
     },
@@ -1703,13 +1703,13 @@ export const models = {
      * (384x384, 512x512, 640x640).
      */
     YOLO26: {
-      ...YOLO26_NANO_SEG_384_XNNPACK_FP32,
+      DEFAULT: YOLO26_NANO_SEG_384_XNNPACK_FP32,
       /**
        * Nano scale YOLO26 instance segmentation model. High speed, ultra low
        * latency mask generation.
        */
       NANO: {
-        ...YOLO26_NANO_SEG_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_NANO_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_NANO_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_NANO_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_NANO_SEG_640_XNNPACK_FP32 },
@@ -1719,7 +1719,7 @@ export const models = {
        * mask accuracy.
        */
       SMALL: {
-        ...YOLO26_SMALL_SEG_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_SMALL_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_SMALL_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_SMALL_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_SMALL_SEG_640_XNNPACK_FP32 },
@@ -1729,7 +1729,7 @@ export const models = {
        * precision for complex multi-object scenes.
        */
       MEDIUM: {
-        ...YOLO26_MEDIUM_SEG_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_MEDIUM_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_640_XNNPACK_FP32 },
@@ -1739,7 +1739,7 @@ export const models = {
        * segmentation variant for demanding visual pipelines.
        */
       LARGE: {
-        ...YOLO26_LARGE_SEG_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_LARGE_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_LARGE_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_LARGE_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_LARGE_SEG_640_XNNPACK_FP32 },
@@ -1749,7 +1749,7 @@ export const models = {
        * segmentation and mask delineation performance.
        */
       XLARGE: {
-        ...YOLO26_XLARGE_SEG_384_XNNPACK_FP32,
+        DEFAULT: YOLO26_XLARGE_SEG_384_XNNPACK_FP32,
         SIZE_384: { XNNPACK_FP32: YOLO26_XLARGE_SEG_384_XNNPACK_FP32 },
         SIZE_512: { XNNPACK_FP32: YOLO26_XLARGE_SEG_512_XNNPACK_FP32 },
         SIZE_640: { XNNPACK_FP32: YOLO26_XLARGE_SEG_640_XNNPACK_FP32 },
@@ -1768,7 +1768,7 @@ export const models = {
      * probability chunks for live mic streaming and STT preprocessing.
      */
     FSMN_VAD: {
-      ...FSMN_VAD_XNNPACK_FP32,
+      DEFAULT: FSMN_VAD_XNNPACK_FP32,
       XNNPACK_FP32: FSMN_VAD_XNNPACK_FP32,
     },
   },
@@ -1788,7 +1788,7 @@ export const models = {
        * speech recognition.
        */
       TINY: {
-        ...WHISPER_TINY_XNNPACK_FP32,
+        DEFAULT: WHISPER_TINY_XNNPACK_FP32,
         XNNPACK_FP32: WHISPER_TINY_XNNPACK_FP32,
         COREML_FP16: WHISPER_TINY_COREML_FP16,
         MLX_BF16: WHISPER_TINY_MLX_BF16,
@@ -1799,7 +1799,7 @@ export const models = {
        * languages.
        */
       BASE: {
-        ...WHISPER_BASE_XNNPACK_FP32,
+        DEFAULT: WHISPER_BASE_XNNPACK_FP32,
         XNNPACK_FP32: WHISPER_BASE_XNNPACK_FP32,
         COREML_FP16: WHISPER_BASE_COREML_FP16,
         MLX_BF16: WHISPER_BASE_MLX_BF16,
@@ -1810,7 +1810,7 @@ export const models = {
        * multi-language audio.
        */
       SMALL: {
-        ...WHISPER_SMALL_XNNPACK_FP32,
+        DEFAULT: WHISPER_SMALL_XNNPACK_FP32,
         XNNPACK_FP32: WHISPER_SMALL_XNNPACK_FP32,
         COREML_FP16: WHISPER_SMALL_COREML_FP16,
         MLX_INT8: WHISPER_SMALL_MLX_INT8,
@@ -1821,7 +1821,7 @@ export const models = {
          * English-only Whisper Tiny model. Fast and compact for English STT.
          */
         TINY: {
-          ...WHISPER_TINY_EN_XNNPACK_FP32,
+          DEFAULT: WHISPER_TINY_EN_XNNPACK_FP32,
           XNNPACK_FP32: WHISPER_TINY_EN_XNNPACK_FP32,
           COREML_FP16: WHISPER_TINY_EN_COREML_FP16,
           MLX_BF16: WHISPER_TINY_EN_MLX_BF16,
@@ -1832,7 +1832,7 @@ export const models = {
          * recognition.
          */
         BASE: {
-          ...WHISPER_BASE_EN_XNNPACK_FP32,
+          DEFAULT: WHISPER_BASE_EN_XNNPACK_FP32,
           XNNPACK_FP32: WHISPER_BASE_EN_XNNPACK_FP32,
           COREML_FP16: WHISPER_BASE_EN_COREML_FP16,
           MLX_BF16: WHISPER_BASE_EN_MLX_BF16,
@@ -1843,7 +1843,7 @@ export const models = {
          * transcription.
          */
         SMALL: {
-          ...WHISPER_SMALL_EN_XNNPACK_FP32,
+          DEFAULT: WHISPER_SMALL_EN_XNNPACK_FP32,
           XNNPACK_FP32: WHISPER_SMALL_EN_XNNPACK_FP32,
           COREML_FP16: WHISPER_SMALL_EN_COREML_FP16,
           MLX_INT8: WHISPER_SMALL_EN_MLX_INT8,
@@ -1873,7 +1873,7 @@ export const models = {
      * on mobile devices.
      */
     LFM2_5_1_2B: {
-      ...LFM2_5_1_2B_XNNPACK_8DA4W,
+      DEFAULT: LFM2_5_1_2B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: LFM2_5_1_2B_XNNPACK_8DA4W,
       XNNPACK_FP16: LFM2_5_1_2B_XNNPACK_FP16,
       MLX_INT4: LFM2_5_1_2B_MLX_INT4,
@@ -1885,7 +1885,7 @@ export const models = {
      * routing, and low-latency chat on resource-constrained edge hardware.
      */
     LFM2_5_350M: {
-      ...LFM2_5_350M_XNNPACK_8DA4W,
+      DEFAULT: LFM2_5_350M_XNNPACK_8DA4W,
       XNNPACK_8DA4W: LFM2_5_350M_XNNPACK_8DA4W,
       XNNPACK_FP16: LFM2_5_350M_XNNPACK_FP16,
       MLX_INT4: LFM2_5_350M_MLX_INT4,
@@ -1897,7 +1897,7 @@ export const models = {
      * UI element inspection, and low-latency multimodal conversational agents.
      */
     LFM2_5_VL_450M: {
-      ...LFM2_5_VL_450M_XNNPACK_8DA4W,
+      DEFAULT: LFM2_5_VL_450M_XNNPACK_8DA4W,
       XNNPACK_8DA4W: LFM2_5_VL_450M_XNNPACK_8DA4W,
       MLX_INT4: LFM2_5_VL_450M_MLX_INT4,
       VULKAN_8DA4W: LFM2_5_VL_450M_VULKAN_8DA4W,
@@ -1909,7 +1909,7 @@ export const models = {
      * precision and reasoning fidelity than the 450M variant.
      */
     LFM2_5_VL_1_6B: {
-      ...LFM2_5_VL_1_6B_XNNPACK_8DA4W,
+      DEFAULT: LFM2_5_VL_1_6B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: LFM2_5_VL_1_6B_XNNPACK_8DA4W,
       VULKAN_8DA4W: LFM2_5_VL_1_6B_VULKAN_8DA4W,
     },
@@ -1920,7 +1920,7 @@ export const models = {
      * high-fidelity bidirectional Polish-English translation.
      */
     BIELIK_V3_1_5B: {
-      ...BIELIK_V3_1_5B_XNNPACK_8DA4W,
+      DEFAULT: BIELIK_V3_1_5B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: BIELIK_V3_1_5B_XNNPACK_8DA4W,
       XNNPACK_FP16: BIELIK_V3_1_5B_XNNPACK_FP16,
     },
@@ -1932,7 +1932,7 @@ export const models = {
      * assistance.
      */
     LLAMA3_2_1B: {
-      ...LLAMA3_2_1B_SPINQUANT,
+      DEFAULT: LLAMA3_2_1B_SPINQUANT,
       XNNPACK_SPINQUANT: LLAMA3_2_1B_SPINQUANT,
       XNNPACK_BF16: LLAMA3_2_1B_BF16,
     },
@@ -1943,7 +1943,7 @@ export const models = {
      * on-device memory profile.
      */
     LLAMA3_2_3B: {
-      ...LLAMA3_2_3B_SPINQUANT,
+      DEFAULT: LLAMA3_2_3B_SPINQUANT,
       XNNPACK_SPINQUANT: LLAMA3_2_3B_SPINQUANT,
       XNNPACK_BF16: LLAMA3_2_3B_BF16,
     },
@@ -1953,7 +1953,7 @@ export const models = {
      * and background processing on low-power devices.
      */
     SMOLLM2_135M: {
-      ...SMOLLM2_135M_8DA4W,
+      DEFAULT: SMOLLM2_135M_8DA4W,
       XNNPACK_8DA4W: SMOLLM2_135M_8DA4W,
       XNNPACK_BF16: SMOLLM2_135M_BF16,
     },
@@ -1964,7 +1964,7 @@ export const models = {
      * simplification, and structured data extraction.
      */
     SMOLLM2_360M: {
-      ...SMOLLM2_360M_8DA4W,
+      DEFAULT: SMOLLM2_360M_8DA4W,
       XNNPACK_8DA4W: SMOLLM2_360M_8DA4W,
       XNNPACK_BF16: SMOLLM2_360M_BF16,
     },
@@ -1975,7 +1975,7 @@ export const models = {
      * 2B-3B models while maintaining fast on-device inference.
      */
     SMOLLM2_1_7B: {
-      ...SMOLLM2_1_7B_8DA4W,
+      DEFAULT: SMOLLM2_1_7B_8DA4W,
       XNNPACK_8DA4W: SMOLLM2_1_7B_8DA4W,
       XNNPACK_BF16: SMOLLM2_1_7B_BF16,
     },
@@ -1986,7 +1986,7 @@ export const models = {
      * tool calling flows.
      */
     HAMMER2_1_0_5B: {
-      ...HAMMER2_1_0_5B_XNNPACK_8DA4W,
+      DEFAULT: HAMMER2_1_0_5B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: HAMMER2_1_0_5B_XNNPACK_8DA4W,
       XNNPACK_BF16: HAMMER2_1_0_5B_XNNPACK_BF16,
     },
@@ -1996,7 +1996,7 @@ export const models = {
      * output generation on edge devices.
      */
     HAMMER2_1_1_5B: {
-      ...HAMMER2_1_1_5B_XNNPACK_8DA4W,
+      DEFAULT: HAMMER2_1_1_5B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: HAMMER2_1_1_5B_XNNPACK_8DA4W,
       XNNPACK_BF16: HAMMER2_1_1_5B_XNNPACK_BF16,
     },
@@ -2007,7 +2007,7 @@ export const models = {
      * autonomous mobile agent pipelines.
      */
     HAMMER2_1_3B: {
-      ...HAMMER2_1_3B_XNNPACK_8DA4W,
+      DEFAULT: HAMMER2_1_3B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: HAMMER2_1_3B_XNNPACK_8DA4W,
       XNNPACK_BF16: HAMMER2_1_3B_XNNPACK_BF16,
     },
@@ -2018,7 +2018,7 @@ export const models = {
      * synthesis, and structured analytical tasks.
      */
     PHI4_MINI: {
-      ...PHI4_MINI_XNNPACK_8DA4W,
+      DEFAULT: PHI4_MINI_XNNPACK_8DA4W,
       XNNPACK_8DA4W: PHI4_MINI_XNNPACK_8DA4W,
       XNNPACK_BF16: PHI4_MINI_XNNPACK_BF16,
     },
@@ -2029,7 +2029,7 @@ export const models = {
      * lightweight conversational assistants on mobile devices.
      */
     QWEN2_5_0_5B: {
-      ...QWEN2_5_0_5B_XNNPACK_8DA4W,
+      DEFAULT: QWEN2_5_0_5B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: QWEN2_5_0_5B_XNNPACK_8DA4W,
       XNNPACK_BF16: QWEN2_5_0_5B_XNNPACK_BF16,
     },
@@ -2040,7 +2040,7 @@ export const models = {
      * cross-lingual translation.
      */
     QWEN2_5_1_5B: {
-      ...QWEN2_5_1_5B_XNNPACK_8DA4W,
+      DEFAULT: QWEN2_5_1_5B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: QWEN2_5_1_5B_XNNPACK_8DA4W,
       XNNPACK_BF16: QWEN2_5_1_5B_XNNPACK_BF16,
     },
@@ -2050,7 +2050,7 @@ export const models = {
      * languages for in-depth text generation and complex multi-turn dialogue.
      */
     QWEN2_5_3B: {
-      ...QWEN2_5_3B_XNNPACK_8DA4W,
+      DEFAULT: QWEN2_5_3B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: QWEN2_5_3B_XNNPACK_8DA4W,
       XNNPACK_BF16: QWEN2_5_3B_XNNPACK_BF16,
     },
@@ -2061,7 +2061,7 @@ export const models = {
      * turn-taking on mobile devices.
      */
     QWEN3_0_6B: {
-      ...QWEN3_0_6B_XNNPACK_8DA4W,
+      DEFAULT: QWEN3_0_6B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: QWEN3_0_6B_XNNPACK_8DA4W,
       XNNPACK_BF16: QWEN3_0_6B_XNNPACK_BF16,
     },
@@ -2071,7 +2071,7 @@ export const models = {
      * proficiency, and conversational fluidity across multiple languages.
      */
     QWEN3_1_7B: {
-      ...QWEN3_1_7B_XNNPACK_8DA4W,
+      DEFAULT: QWEN3_1_7B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: QWEN3_1_7B_XNNPACK_8DA4W,
       XNNPACK_BF16: QWEN3_1_7B_XNNPACK_BF16,
     },
@@ -2082,7 +2082,7 @@ export const models = {
      * on-device AI applications.
      */
     QWEN3_4B: {
-      ...QWEN3_4B_XNNPACK_8DA4W,
+      DEFAULT: QWEN3_4B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: QWEN3_4B_XNNPACK_8DA4W,
       XNNPACK_BF16: QWEN3_4B_XNNPACK_BF16,
     },
@@ -2093,7 +2093,7 @@ export const models = {
      * for mobile deployment.
      */
     GEMMA4_E2B: {
-      ...GEMMA4_E2B_XNNPACK_8DA4W,
+      DEFAULT: GEMMA4_E2B_XNNPACK_8DA4W,
       XNNPACK_8DA4W: GEMMA4_E2B_XNNPACK_8DA4W,
       MLX_INT4: GEMMA4_E2B_MLX_INT4,
     },
@@ -2110,7 +2110,7 @@ export const models = {
      * sentence similarity, and clustering.
      */
     ALL_MINILM_L6_V2: {
-      ...ALL_MINILM_L6_V2_EMBEDDINGS,
+      DEFAULT: ALL_MINILM_L6_V2_EMBEDDINGS,
       XNNPACK_FP32: ALL_MINILM_L6_V2_EMBEDDINGS,
     },
     /**
@@ -2118,7 +2118,7 @@ export const models = {
      * Provides higher quality semantic embeddings compared to MiniLM.
      */
     ALL_MPNET_BASE_V2: {
-      ...ALL_MPNET_BASE_V2_EMBEDDINGS,
+      DEFAULT: ALL_MPNET_BASE_V2_EMBEDDINGS,
       XNNPACK_FP32: ALL_MPNET_BASE_V2_EMBEDDINGS,
     },
     /**
@@ -2126,7 +2126,7 @@ export const models = {
      * QA matching using cosine similarity.
      */
     MULTI_QA_MINILM_L6_COS_V1: {
-      ...MULTI_QA_MINILM_L6_COS_V1_EMBEDDINGS,
+      DEFAULT: MULTI_QA_MINILM_L6_COS_V1_EMBEDDINGS,
       XNNPACK_FP32: MULTI_QA_MINILM_L6_COS_V1_EMBEDDINGS,
     },
     /**
@@ -2134,7 +2134,7 @@ export const models = {
      * question-answering matching using dot product distance.
      */
     MULTI_QA_MPNET_BASE_DOT_V1: {
-      ...MULTI_QA_MPNET_BASE_DOT_V1_EMBEDDINGS,
+      DEFAULT: MULTI_QA_MPNET_BASE_DOT_V1_EMBEDDINGS,
       XNNPACK_FP32: MULTI_QA_MPNET_BASE_DOT_V1_EMBEDDINGS,
     },
     /**
@@ -2142,7 +2142,7 @@ export const models = {
      * cross-lingual semantic similarity.
      */
     PARAPHRASE_MULTILINGUAL_MINILM_L12_V2: {
-      ...PARAPHRASE_MULTILINGUAL_MINILM_L12_V2_EMBEDDINGS,
+      DEFAULT: PARAPHRASE_MULTILINGUAL_MINILM_L12_V2_EMBEDDINGS,
       XNNPACK_8DA4W: PARAPHRASE_MULTILINGUAL_MINILM_L12_V2_EMBEDDINGS,
     },
     /**
@@ -2150,7 +2150,7 @@ export const models = {
      * distilled Universal Sentence Encoder (512-dim output).
      */
     DISTILUSE_BASE_MULTILINGUAL_CASED_V2: {
-      ...DISTILUSE_BASE_MULTILINGUAL_CASED_V2_EMBEDDINGS,
+      DEFAULT: DISTILUSE_BASE_MULTILINGUAL_CASED_V2_EMBEDDINGS,
       XNNPACK_8DA4W: DISTILUSE_BASE_MULTILINGUAL_CASED_V2_EMBEDDINGS,
       MLX_INT8: DISTILUSE_BASE_MULTILINGUAL_CASED_V2_MLX_INT8,
     },
@@ -2161,7 +2161,7 @@ export const models = {
      * search.
      */
     CLIP_VIT_BASE_PATCH32_TEXT: {
-      ...CLIP_VIT_BASE_PATCH32_TEXT_EMBEDDINGS,
+      DEFAULT: CLIP_VIT_BASE_PATCH32_TEXT_EMBEDDINGS,
       XNNPACK_FP32: CLIP_VIT_BASE_PATCH32_TEXT_EMBEDDINGS,
     },
     /**
@@ -2170,7 +2170,7 @@ export const models = {
      * passages with `document: ` via `embed(text, 'document: ')`.
      */
     LFM2_5_EMBEDDING_350M: {
-      ...LFM2_5_EMBEDDING_350M_EMBEDDINGS,
+      DEFAULT: LFM2_5_EMBEDDING_350M_EMBEDDINGS,
       XNNPACK_8DA4W: LFM2_5_EMBEDDING_350M_EMBEDDINGS,
       MLX_INT4: LFM2_5_EMBEDDING_350M_MLX_INT4,
     },
@@ -2187,7 +2187,7 @@ export const models = {
      * address, and similar). Compact label space, best for general redaction.
      */
     OPENAI: {
-      ...PRIVACY_FILTER_OPENAI_XNNPACK_8DA4W,
+      DEFAULT: PRIVACY_FILTER_OPENAI_XNNPACK_8DA4W,
       XNNPACK_8DA4W: PRIVACY_FILTER_OPENAI_XNNPACK_8DA4W,
       MLX_INT4: PRIVACY_FILTER_OPENAI_MLX_INT4,
     },
@@ -2196,7 +2196,7 @@ export const models = {
      * space for stricter compliance-oriented redaction.
      */
     NEMOTRON: {
-      ...PRIVACY_FILTER_NEMOTRON_XNNPACK_8DA4W,
+      DEFAULT: PRIVACY_FILTER_NEMOTRON_XNNPACK_8DA4W,
       XNNPACK_8DA4W: PRIVACY_FILTER_NEMOTRON_XNNPACK_8DA4W,
       MLX_INT8: PRIVACY_FILTER_NEMOTRON_MLX_INT8,
     },
@@ -2212,7 +2212,7 @@ export const models = {
      * cross-modal image search.
      */
     CLIP_VIT_BASE_PATCH32: {
-      ...CLIP_VIT_BASE_PATCH32_IMAGE_XNNPACK_FP32,
+      DEFAULT: CLIP_VIT_BASE_PATCH32_IMAGE_XNNPACK_FP32,
       XNNPACK_FP32: CLIP_VIT_BASE_PATCH32_IMAGE_XNNPACK_FP32,
       COREML_FP16: CLIP_VIT_BASE_PATCH32_IMAGE_COREML_FP16,
       MLX_INT8: CLIP_VIT_BASE_PATCH32_IMAGE_MLX_INT8,
@@ -2229,7 +2229,7 @@ export const models = {
      * text prompts in real time.
      */
     SDXS_512_DREAMSHAPER: {
-      ...SDXS_512_DREAMSHAPER_XNNPACK_FP32,
+      DEFAULT: SDXS_512_DREAMSHAPER_XNNPACK_FP32,
       XNNPACK_FP32: SDXS_512_DREAMSHAPER_XNNPACK_FP32,
       COREML_FP16: SDXS_512_DREAMSHAPER_COREML_FP16,
     },
@@ -2246,7 +2246,7 @@ export const models = {
      * voice presets (see {@link SUPERTONIC_DEFAULT_VOICE_NAMES}).
      */
     SUPERTONIC: {
-      ...SUPERTONIC_3_XNNPACK_FP32,
+      DEFAULT: SUPERTONIC_3_XNNPACK_FP32,
       XNNPACK_FP32: SUPERTONIC_3_XNNPACK_FP32,
       MLX_FP32: SUPERTONIC_3_MLX_FP32,
     },
@@ -2258,39 +2258,39 @@ export const models = {
      */
     KOKORO: {
       EN_US: {
-        ...KOKORO_EN_US_XNNPACK_FP32,
+        DEFAULT: KOKORO_EN_US_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_EN_US_XNNPACK_FP32,
       },
       EN_GB: {
-        ...KOKORO_EN_GB_XNNPACK_FP32,
+        DEFAULT: KOKORO_EN_GB_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_EN_GB_XNNPACK_FP32,
       },
       ES: {
-        ...KOKORO_ES_XNNPACK_FP32,
+        DEFAULT: KOKORO_ES_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_ES_XNNPACK_FP32,
       },
       FR: {
-        ...KOKORO_FR_XNNPACK_FP32,
+        DEFAULT: KOKORO_FR_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_FR_XNNPACK_FP32,
       },
       IT: {
-        ...KOKORO_IT_XNNPACK_FP32,
+        DEFAULT: KOKORO_IT_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_IT_XNNPACK_FP32,
       },
       PT: {
-        ...KOKORO_PT_XNNPACK_FP32,
+        DEFAULT: KOKORO_PT_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_PT_XNNPACK_FP32,
       },
       HI: {
-        ...KOKORO_HI_XNNPACK_FP32,
+        DEFAULT: KOKORO_HI_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_HI_XNNPACK_FP32,
       },
       PL: {
-        ...KOKORO_PL_XNNPACK_FP32,
+        DEFAULT: KOKORO_PL_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_PL_XNNPACK_FP32,
       },
       DE: {
-        ...KOKORO_DE_XNNPACK_FP32,
+        DEFAULT: KOKORO_DE_XNNPACK_FP32,
         XNNPACK_FP32: KOKORO_DE_XNNPACK_FP32,
       },
     },
@@ -2312,7 +2312,7 @@ export const models = {
        * PP-OCRv6 Small multilingual OCR model.
        */
       PPOCRV6_SMALL: {
-        ...PPOCRV6_SMALL_XNNPACK_INT8,
+        DEFAULT: PPOCRV6_SMALL_XNNPACK_INT8,
         XNNPACK: PPOCRV6_SMALL_XNNPACK_INT8,
         COREML: PPOCRV6_SMALL_COREML_INT8,
         VULKAN: PPOCRV6_SMALL_VULKAN_FP16,
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index 3298a5a15d..aed28dfe95 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -1,7 +1,6 @@
 /**
  * Utility functions for inspecting models and querying native runtime
  * capabilities.
- * @module Utils
  */
 
 import { rnexecutorchJsi } from './native/bridge';
@@ -12,7 +11,7 @@ import RNBlobUtil from 'react-native-blob-util';
 /**
  * Retrieves the names of all ExecuTorch backends compiled and registered in the
  * native binary.
- * @category Core / Functions
+ * @category Utils / Functions
  * @returns An array of registered backend name strings (e.g. 'XnnpackBackend',
  * 'CoreMLBackend').
  */
@@ -21,9 +20,25 @@ export function getRegisteredBackends(): string[] {
   return rnexecutorchJsi.getExecuTorchRegisteredBackends();
 }
 
+/**
+ * Options accepted by {@link useResourceDownload} and by every `use<Task>` hook
+ * built on top of it.
+ * @category Utils / Types
+ */
+export type ResourceOptions = {
+  /** If true, prevents checks and downloads, resetting the hook state. */
+  preventLoad?: boolean;
+  /**
+   * Re-downloads every remote source even when it is already cached, replacing
+   * the cached copy. Use to recover from a corrupted file or to pick up a model
+   * that changed behind a stable URL.
+   */
+  forceDownload?: boolean;
+};
+
 /**
  * Result of inspecting an ExecuTorch model file.
- * @category Core / Types
+ * @category Utils / Types
  */
 export type ModelInspection = {
   /** The model source URL or file path that was inspected. */
@@ -48,7 +63,7 @@ export type ModelInspection = {
  * is not reused. Inspecting a remote model therefore re-downloads it on every
  * call and leaves nothing behind — call {@link download} first and inspect the
  * returned local path if you also intend to run the model.
- * @category Core / Functions
+ * @category Utils / Functions
  * @param source The remote HTTP URL or local path to the `.pte` model file.
  * @returns A promise resolving to an object containing the model source, method
  * signature metadata, and per-method backend usage.

From 449f1f5cc69154573ad36f4def8da2c6d61d7ccb Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 14:45:10 +0200
Subject: [PATCH 21/38] docs: link create<Task> return types and expose flat CV
 operations

---
 .../react-native-executorch/src/extensions/cv/ops/index.ts   | 5 +++++
 .../src/extensions/cv/tasks/classification.ts                | 3 +--
 .../src/extensions/cv/tasks/imageEmbedding.ts                | 3 +--
 .../src/extensions/cv/tasks/instanceSegmentation.ts          | 3 +--
 .../src/extensions/cv/tasks/keypointDetection.ts             | 3 +--
 .../src/extensions/cv/tasks/objectDetection.ts               | 3 +--
 .../src/extensions/cv/tasks/paddleOcr.ts                     | 3 +--
 .../src/extensions/cv/tasks/preprocessing.ts                 | 3 +--
 .../src/extensions/cv/tasks/sdxsTextToImage.ts               | 2 +-
 .../src/extensions/cv/tasks/semanticSegmentation.ts          | 3 +--
 .../src/extensions/cv/tasks/styleTransfer.ts                 | 2 +-
 .../src/extensions/llm/tasks/llmChatSession.ts               | 2 +-
 .../src/extensions/nlp/tasks/privacyFilter.ts                | 3 +--
 .../src/extensions/nlp/tasks/textEmbedding.ts                | 3 +--
 .../src/extensions/nlp/tasks/tokenization.ts                 | 4 ++--
 .../extensions/speech/tasks/fsmnVoiceActivityDetection.ts    | 3 +--
 .../src/extensions/speech/tasks/kokoroTextToSpeech.ts        | 2 +-
 .../src/extensions/speech/tasks/supertonicTextToSpeech.ts    | 3 ++-
 .../src/extensions/speech/tasks/whisperSpeechToText.ts       | 4 ++--
 19 files changed, 26 insertions(+), 31 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/ops/index.ts b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
index 04a028b2fb..e310574d16 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
@@ -1,6 +1,11 @@
 /**
  * Aggregated image, box, and point operations.
  */
+export * from './image';
+export * from './boxes';
+export * from './points';
+export * from './quad';
+
 export * as image from './image';
 export * as boxes from './boxes';
 export * as points from './points';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
index f2e790ba17..efbc2cf36b 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
@@ -111,8 +111,7 @@ export type Classifier<L> = {
  * See {@link ClassifierModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing classification and
- * disposal controls.
+ * @returns A promise resolving to the instantiated {@link Classifier} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
  * `SCHEMA_MISMATCH` if model schema does not match classification spec, or
  * `INVALID_ARGUMENT` if labels length does not match model output classes.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
index 5e26ff6329..a2e95f07ff 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
@@ -67,8 +67,7 @@ export type ImageEmbedder = {
  * See {@link ImageEmbedderModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing the embedding and
- * disposal controls.
+ * @returns A promise resolving to the instantiated {@link ImageEmbedder} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
  * or `SCHEMA_MISMATCH` if model schema does not match embedding spec.
  */
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index 96379a19dc..df3c4a63a3 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -159,8 +159,7 @@ export type InstanceSegmenter<F extends BoxFormat, L> = {
  * See {@link InstanceSegmenterModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing instance segmentation
- * and disposal controls.
+ * @returns A promise resolving to the instantiated {@link InstanceSegmenter} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
  * or `SCHEMA_MISMATCH` if model schema does not match instance segmentation
  * spec.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index 54ebd50142..9cd2398d53 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -216,8 +216,7 @@ function postprocess<F extends BoxFormat, L extends PropertyKey>(
  * See {@link KeypointDetectorModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing keypoint detection and
- * disposal bindings.
+ * @returns A promise resolving to the instantiated {@link KeypointDetector} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
  * or `SCHEMA_MISMATCH` if model schema does not match keypoint detection spec.
  */
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index 70c0d7ae3f..2c65bb5f63 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -135,8 +135,7 @@ export type ObjectDetector<F extends BoxFormat, L> = {
  * See {@link ObjectDetectorModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing object detection and
- * disposal controls.
+ * @returns A promise resolving to the instantiated {@link ObjectDetector} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
  * or `SCHEMA_MISMATCH` if model schema does not match object detection spec.
  */
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 9c7ad3389b..7e1c4f0ec3 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -390,8 +390,7 @@ function greedyCtcDecode(
  * and detection thresholds. See {@link PaddleOcrModel}.
  * @param runtime Optional worklet runtime thread on which to run detection and
  * recognition.
- * @returns A promise resolving to an object containing recognition and disposal
- * controls.
+ * @returns A promise resolving to the instantiated {@link PaddleOcr} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model or charset
  * fails to load, `SCHEMA_MISMATCH` if the loaded model does not match the
  * PP-OCRv6 detect/recognize contract, or if the charset does not match the
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
index 203d676760..6970047fd9 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
@@ -70,8 +70,7 @@ export type ImagePreprocessor = {
  * See {@link ImagePreprocessorOptions}.
  * @param outputShape Expected output shape of the preprocessed model input
  * tensor (must match rank-3 `[3, H, W]` or rank-4 `[1, 3, H, W]`).
- * @returns An object containing the `process` runner function and a `dispose`
- * method.
+ * @returns An instantiated {@link ImagePreprocessor} pipeline.
  * @throws {RnExecuTorchError} With code `SCHEMA_MISMATCH` if `outputShape` does
  * not match rank-3 `[3, H, W]` or rank-4 `[1, 3, H, W]`.
  */
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
index 8b06f5ffbc..37512c737d 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/sdxsTextToImage.ts
@@ -86,7 +86,7 @@ export type SdxsTextToImage = {
  * @param config SDXS pipeline configuration containing the model and tokenizer paths.
  * See {@link SdxsTextToImageModel}.
  * @param runtime Optional worklet runtime thread on which to run generation.
- * @returns A promise resolving to an object with generation and disposal controls.
+ * @returns A promise resolving to the instantiated {@link SdxsTextToImage} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model or tokenizer
  * fails to load, or `SCHEMA_MISMATCH` if model schema does not match SDXS spec.
  */
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
index 2d4b8d3867..dea0ec5c0e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
@@ -140,8 +140,7 @@ function hslToRgb(h: number, s: number, l: number): [number, number, number] {
  * @param config Segmenter task configuration containing path and options.
  * See {@link SemanticSegmenterModel}.
  * @param runtime Optional worklet runtime thread environment context.
- * @returns A promise resolving to an object containing segmentation and
- * disposal controls.
+ * @returns A promise resolving to the instantiated {@link SemanticSegmenter} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
  * `SCHEMA_MISMATCH` if model schema does not match segmentation spec, or
  * `INVALID_ARGUMENT` if labels length does not match model output classes.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
index c14a8e0ef3..8a537ad87a 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
@@ -85,7 +85,7 @@ export type StyleTransfer = {
  * @param config Style transfer task configuration containing path and options.
  * See {@link StyleTransferModel}.
  * @param runtime Optional worklet runtime thread on which to run the model execution.
- * @returns A promise resolving to an object containing style transfer and disposal controls.
+ * @returns A promise resolving to the instantiated {@link StyleTransfer} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if model fails to load,
  * or `SCHEMA_MISMATCH` if model schema does not match style transfer spec.
  */
diff --git a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
index c1a1d7b24a..cfbb5a42d1 100644
--- a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
+++ b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
@@ -190,7 +190,7 @@ const DEFAULT_MAX_TURNS = 5;
  * @param config Model configuration containing model, tokenizer, and tokenizer config paths.
  * @param options Custom generation, tool calling, and state options.
  * @param runtime The worklet runtime thread to run native generation on.
- * @returns A Promise resolving to an {@link LLMChatSession} instance.
+ * @returns A promise resolving to the instantiated {@link LLMChatSession} session.
  */
 export async function createLLMChatSession(
   config: LLMModel,
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
index 2586840ccc..e22dd9861f 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
@@ -118,8 +118,7 @@ export type PrivacyFilter<Label extends string = string> = {
  * tokenizer paths plus the label space options. See {@link PrivacyFilterModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing detection and disposal
- * controls.
+ * @returns A promise resolving to the instantiated {@link PrivacyFilter} runner.
  * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if `labelNames` is
  * empty or does not start with `'O'`, `LOAD_FAILED` if the model or tokenizer
  * fails to load, or `SCHEMA_MISMATCH` if the loaded model schema does not match
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
index bb7f5124a4..d5c828b1af 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
@@ -71,8 +71,7 @@ export type TextEmbedder = {
  * tokenizer paths. See {@link TextEmbedderModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing the embedding and
- * disposal controls.
+ * @returns A promise resolving to the instantiated {@link TextEmbedder} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model or tokenizer
  * fails to load, or `SCHEMA_MISMATCH` if the model schema does not match the
  * text embedding specification.
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
index bb506fe232..722e3e0195 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
@@ -13,8 +13,8 @@ import { loadTokenizer } from '../tokenizer';
  * @category NLP / Tasks
  * @param tokenizerPath Absolute local path to a `tokenizer.json` file.
  * @param runtime Optional worklet runtime thread to run the tokenizer on.
- * @returns A promise resolving to the tokenizer operations and a `dispose`
- * handle that releases the native tokenizer.
+ * @returns A promise resolving to the instantiated {@link Tokenizer} operations
+ * and disposal controls.
  */
 export async function createTokenizer(tokenizerPath: string, runtime?: WorkletRuntime) {
   const tokenizer = await wrapAsync(loadTokenizer, runtime)(tokenizerPath);
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
index 4c923e2981..242b56801f 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/fsmnVoiceActivityDetection.ts
@@ -236,8 +236,7 @@ export type FsmnVoiceActivityDetector = {
  * options. See {@link FsmnVadModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing detection and disposal
- * controls.
+ * @returns A promise resolving to the instantiated {@link FsmnVoiceActivityDetector} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model fails to
  * load, or `SCHEMA_MISMATCH` if the model schema does not match the VAD
  * specification.
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
index 7cfbcaad8c..70059a214c 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
@@ -145,7 +145,7 @@ export type KokoroTextToSpeech<K extends PropertyKey = string> = {
  * @param config Kokoro TTS pipeline configuration containing model and asset paths.
  * See {@link KokoroTtsModel}.
  * @param runtime Optional worklet runtime thread on which to run inference.
- * @returns A promise resolving to an object with audio synthesis and disposal controls.
+ * @returns A promise resolving to the instantiated {@link KokoroTextToSpeech} runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if models or voice files
  * fail to load, `SCHEMA_MISMATCH` if model schemas do not match the Kokoro
  * specification, or `INVALID_STATE` if phonemizer native support is missing.
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
index 68d14c5244..62129979ee 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
@@ -166,7 +166,8 @@ export type SupertonicTextToSpeech<K extends PropertyKey = string> = {
  * @param config Supertonic TTS pipeline configuration containing model and asset paths.
  * See {@link SupertonicTtsModel}.
  * @param runtime Optional worklet runtime thread on which to run inference.
- * @returns A promise resolving to an object with audio synthesis and disposal controls.
+ * @returns A promise resolving to the instantiated {@link SupertonicTextToSpeech}
+ * runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if models, indexer, or
  * voice styles fail to load, or `SCHEMA_MISMATCH` if model schemas do not match
  * the Supertonic specification.
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
index 4fd3f6458e..dc25cceef4 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts
@@ -185,8 +185,8 @@ export type WhisperSpeechToText<L extends WhisperLanguage = WhisperLanguage> = {
  * See {@link WhisperSttModel}.
  * @param runtime Optional worklet runtime thread on which to run the model
  * execution.
- * @returns A promise resolving to an object containing transcription and
- * disposal controls.
+ * @returns A promise resolving to the instantiated {@link WhisperSpeechToText}
+ * runner.
  * @throws {RnExecuTorchError} With code `LOAD_FAILED` if the model, tokenizer,
  * or VAD fails to load, or `SCHEMA_MISMATCH` if model schemas do not match the
  * Whisper specification.

From 3cdd04aacf2a5bc7718421d8d3aa68cc7f73a6c7 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:00:41 +0200
Subject: [PATCH 22/38] refactor: flatten cv exports, add package subpaths, and
 eliminate re-export leaks

---
 apps/computer-vision/app/inspect/index.tsx    |  6 ++--
 apps/nlp/app/llm/index.tsx                    | 19 ++++-------
 apps/nlp/app/privacy-filter/index.tsx         | 15 +++-----
 apps/nlp/constants/llm.ts                     | 34 +++++++------------
 apps/speech/app/text-to-speech/index.tsx      |  7 ++--
 docs/typedoc.json                             |  7 ----
 packages/react-native-executorch/package.json | 30 ++++++++++++++++
 .../src/extensions/cv/ops/boxes.ts            |  9 -----
 .../src/extensions/cv/ops/image.ts            | 10 ------
 .../src/extensions/cv/ops/index.ts            |  7 +---
 .../src/extensions/cv/ops/points.ts           |  5 ---
 .../src/extensions/cv/ops/quad.ts             |  8 -----
 .../cv/tasks/instanceSegmentation.ts          |  2 --
 .../extensions/cv/tasks/keypointDetection.ts  |  2 --
 .../extensions/cv/tasks/objectDetection.ts    |  2 --
 .../src/extensions/cv/utils/paddleOcrUtils.ts |  2 --
 .../extensions/llm/tasks/llmChatSession.ts    | 22 +-----------
 .../src/extensions/nlp/tasks/privacyFilter.ts |  5 ---
 .../speech/tasks/supertonicTextToSpeech.ts    |  3 --
 .../src/hooks/useInstanceSegmenter.ts         |  2 +-
 .../src/hooks/useKeypointDetector.ts          |  2 +-
 .../src/hooks/useObjectDetector.ts            |  2 +-
 packages/react-native-executorch/src/index.ts |  2 --
 23 files changed, 64 insertions(+), 139 deletions(-)

diff --git a/apps/computer-vision/app/inspect/index.tsx b/apps/computer-vision/app/inspect/index.tsx
index c22bb5a832..9620063905 100644
--- a/apps/computer-vision/app/inspect/index.tsx
+++ b/apps/computer-vision/app/inspect/index.tsx
@@ -10,13 +10,13 @@ import {
   Alert,
 } from 'react-native';
 import { useSafeAreaInsets } from 'react-native-safe-area-context';
-import { inspectModel, type ConcreteDim, type ParamSpec } from 'react-native-executorch';
+import { inspectModel, type schema } from 'react-native-executorch';
 import ScreenWrapper from '../../components/ScreenWrapper';
 import { ColorPalette } from '../../theme';
 
 type InspectionResult = Awaited<ReturnType<typeof inspectModel>>;
 
-const formatDim = (dim: ConcreteDim): string => {
+const formatDim = (dim: schema.ConcreteDim): string => {
   switch (dim.kind) {
     case 'constant':
       return `${dim.value}`;
@@ -55,7 +55,7 @@ function InspectContent() {
   };
 
   const renderParamList = (
-    params: readonly ParamSpec<ConcreteDim>[] | undefined,
+    params: readonly schema.ParamSpec<schema.ConcreteDim>[] | undefined,
     title: string
   ) => {
     if (!params || params.length === 0) return null;
diff --git a/apps/nlp/app/llm/index.tsx b/apps/nlp/app/llm/index.tsx
index 3955ff2d62..658435e62a 100644
--- a/apps/nlp/app/llm/index.tsx
+++ b/apps/nlp/app/llm/index.tsx
@@ -15,14 +15,7 @@ import {
 import { Skia } from '@shopify/react-native-skia';
 import RNBlobUtil from 'react-native-blob-util';
 import { useSafeAreaInsets } from 'react-native-safe-area-context';
-import {
-  useLLMChatSession,
-  type LLMGenerationStats,
-  type LLMKVCacheState,
-  type ToolCall,
-  type ChatMessage,
-  type cv,
-} from 'react-native-executorch';
+import { useLLMChatSession, type llm, type cv } from 'react-native-executorch';
 import ScreenWrapper from '../../components/ScreenWrapper';
 import { ModelPicker, type ModelOption } from '../../components/ModelPicker';
 import { Button } from '../../components/Button';
@@ -33,11 +26,11 @@ type Turn = {
   role: 'user' | 'assistant' | 'tool';
   content: string;
   imageUri?: string;
-  stats?: LLMGenerationStats;
-  toolCalls?: readonly ToolCall[];
+  stats?: llm.LLMGenerationStats;
+  toolCalls?: readonly llm.ToolCall[];
 };
 
-function formatStats(stats: LLMGenerationStats): string {
+function formatStats(stats: llm.LLMGenerationStats): string {
   const decodeMs = stats.inferenceEndMs - stats.firstTokenMs;
   const tokensPerSec = decodeMs > 0 ? (stats.numGeneratedTokens / decodeMs) * 1000 : 0;
   const decodeTtftMs = stats.firstTokenMs - stats.inferenceStartMs;
@@ -70,7 +63,7 @@ function LLMContent() {
     []
   );
 
-  const initialMessages: ChatMessage[] = useMemo(
+  const initialMessages: llm.ChatMessage[] = useMemo(
     () => (activeModel.systemPrompt ? [{ role: 'system', content: activeModel.systemPrompt }] : []),
     [activeModel]
   );
@@ -88,7 +81,7 @@ function LLMContent() {
   const [turns, setTurns] = useState<Turn[]>([]);
   const [streamingResponse, setStreamingResponse] = useState<string | null>(null);
 
-  let kvCacheState: LLMKVCacheState | null = null;
+  let kvCacheState: llm.LLMKVCacheState | null = null;
   if (isReady && getKVCacheState) {
     try {
       kvCacheState = getKVCacheState();
diff --git a/apps/nlp/app/privacy-filter/index.tsx b/apps/nlp/app/privacy-filter/index.tsx
index 37ea7e8189..e5aed1838d 100644
--- a/apps/nlp/app/privacy-filter/index.tsx
+++ b/apps/nlp/app/privacy-filter/index.tsx
@@ -9,14 +9,7 @@ import {
   KeyboardAvoidingView,
   Platform,
 } from 'react-native';
-import {
-  usePrivacyFilter,
-  models,
-  piiSegments,
-  type PiiEntity,
-  type PiiSegment,
-  type PrivacyFilterModel,
-} from 'react-native-executorch';
+import { usePrivacyFilter, models, nlp, type PrivacyFilterModel } from 'react-native-executorch';
 import ScreenWrapper from '../../components/ScreenWrapper';
 import { ModelStatus } from '../../components/ModelStatus';
 import { Button } from '../../components/Button';
@@ -68,14 +61,14 @@ function PrivacyFilterContent() {
   const { isReady, downloadProgress, error, detectPii } = usePrivacyFilter(active.value);
 
   const [text, setText] = useState(active.sample);
-  const [entities, setEntities] = useState<PiiEntity[] | null>(null);
+  const [entities, setEntities] = useState<nlp.PiiEntity[] | null>(null);
   const [busy, setBusy] = useState(false);
   const [runError, setRunError] = useState<string | null>(null);
   const [inferenceMs, setInferenceMs] = useState<number | null>(null);
 
   const ready = isReady && !!detectPii;
-  const segments: PiiSegment[] | null = useMemo(
-    () => (entities ? piiSegments(text, entities) : null),
+  const segments: nlp.PiiSegment[] | null = useMemo(
+    () => (entities ? nlp.piiSegments(text, entities) : null),
     [text, entities]
   );
 
diff --git a/apps/nlp/constants/llm.ts b/apps/nlp/constants/llm.ts
index 19f7892991..405e531a25 100644
--- a/apps/nlp/constants/llm.ts
+++ b/apps/nlp/constants/llm.ts
@@ -1,14 +1,6 @@
-import {
-  models,
-  type LLMModel,
-  type LLMToolOpts,
-  type LLMGenerationConfig,
-  type ToolDefinition,
-  type ToolParserResult,
-  type ToolCall,
-} from 'react-native-executorch';
-
-export const TOOLS: ToolDefinition[] = [
+import { models, type llm, type LLMModel, type LLMToolOpts } from 'react-native-executorch';
+
+export const TOOLS: llm.ToolDefinition[] = [
   {
     type: 'function',
     function: {
@@ -139,10 +131,10 @@ export const TOOLS: ToolDefinition[] = [
 
 export const GEMMA_TOOL_STOP_REGEX = /(?:<tool_call\|>|<turn\|>)/;
 
-export function parseGemmaToolCalls(text: string): ToolParserResult | undefined {
+export function parseGemmaToolCalls(text: string): llm.ToolParserResult | undefined {
   const callRegex = /<\|tool_call>call:([a-zA-Z0-9_-]+)\{([\s\S]*?)\}(?:<tool_call\|>)?/g;
 
-  const toolCalls: ToolCall[] = [];
+  const toolCalls: llm.ToolCall[] = [];
   for (const match of text.matchAll(callRegex)) {
     const name = match[1]!;
     const rawArgs = match[2]?.trim() ?? '';
@@ -173,7 +165,7 @@ export function parseGemmaToolCalls(text: string): ToolParserResult | undefined
 
 export const HAMMER_TOOL_STOP_REGEX = /(?:<\|im_end\|>)/;
 
-export function parseHammerToolCalls(text: string): ToolParserResult | undefined {
+export function parseHammerToolCalls(text: string): llm.ToolParserResult | undefined {
   const jsonMatch =
     text.match(/```(?:json)?\s*([\s\S]*?)\s*```/) ?? text.match(/\[\s*\{[\s\S]*\}\s*\]/);
   const jsonStr = jsonMatch ? (jsonMatch[1] ?? jsonMatch[0]) : text.trim();
@@ -182,7 +174,7 @@ export function parseHammerToolCalls(text: string): ToolParserResult | undefined
     const parsed = JSON.parse(jsonStr);
     if (!Array.isArray(parsed) || parsed.length === 0) return undefined;
 
-    const toolCalls: ToolCall[] = [];
+    const toolCalls: llm.ToolCall[] = [];
     for (const item of parsed) {
       if (item && typeof item === 'object' && typeof item.name === 'string') {
         toolCalls.push({
@@ -205,9 +197,9 @@ export function parseHammerToolCalls(text: string): ToolParserResult | undefined
 
 export const QWEN_TOOL_STOP_REGEX = /(?:<\/tool_call>|<\|im_end\|>)/;
 
-export function parseQwenToolCalls(text: string): ToolParserResult | undefined {
+export function parseQwenToolCalls(text: string): llm.ToolParserResult | undefined {
   const callRegex = /<tool_call>\s*(\{[\s\S]*?\})\s*<\/tool_call>/g;
-  const toolCalls: ToolCall[] = [];
+  const toolCalls: llm.ToolCall[] = [];
   let match;
 
   while ((match = callRegex.exec(text)) !== null) {
@@ -245,7 +237,7 @@ export function parseQwenToolCalls(text: string): ToolParserResult | undefined {
 
 export const LLAMA_TOOL_STOP_REGEX = /(?:<\|eot_id\|>|<\|eom_id\|>)/;
 
-export function parseLlamaToolCalls(text: string): ToolParserResult | undefined {
+export function parseLlamaToolCalls(text: string): llm.ToolParserResult | undefined {
   const cleanText = text
     .replace(/<\|(?:eot_id|eom_id|start_header_id|end_header_id)\|>[\s\S]*?$/g, '')
     .trim();
@@ -270,7 +262,7 @@ export function parseLlamaToolCalls(text: string): ToolParserResult | undefined
           args = {};
         }
       }
-      const toolCalls: ToolCall[] = [
+      const toolCalls: llm.ToolCall[] = [
         {
           type: 'function',
           function: {
@@ -293,7 +285,7 @@ export function parseLlamaToolCalls(text: string): ToolParserResult | undefined
 
 // --- Common Configurations ---
 
-export const DEFAULT_GENERATION_CONFIG: LLMGenerationConfig = {
+export const DEFAULT_GENERATION_CONFIG: llm.LLMGenerationConfig = {
   temperature: 0.7,
   maxNewTokens: 512,
   echo: false,
@@ -327,7 +319,7 @@ export interface LLMModelConfig {
   id: string;
   name: string;
   model: LLMModel;
-  generationConfig: LLMGenerationConfig;
+  generationConfig: llm.LLMGenerationConfig;
   systemPrompt?: string;
   stopRegex?: RegExp;
   toolOpts: LLMToolOpts | undefined;
diff --git a/apps/speech/app/text-to-speech/index.tsx b/apps/speech/app/text-to-speech/index.tsx
index ecdb1fcccc..b6dcb2be1b 100644
--- a/apps/speech/app/text-to-speech/index.tsx
+++ b/apps/speech/app/text-to-speech/index.tsx
@@ -3,11 +3,10 @@ import { Platform, View, Text, StyleSheet, ScrollView, TextInput } from 'react-n
 import {
   useTextToSpeech,
   models,
+  speech,
   SUPERTONIC_SAMPLE_RATE,
-  SUPERTONIC_SUPPORTED_LANGUAGES,
   SUPERTONIC_DEFAULT_VOICE_NAMES,
   type SupertonicDefaultVoiceName,
-  type SupertonicLanguage,
 } from 'react-native-executorch';
 import { AudioContext, type AudioBufferQueueSourceNode } from 'react-native-audio-api';
 
@@ -30,7 +29,7 @@ const VOICE_OPTIONS = SUPERTONIC_DEFAULT_VOICE_NAMES.map((name) => ({
   value: name as SupertonicDefaultVoiceName,
 }));
 
-const LANGUAGE_OPTIONS = SUPERTONIC_SUPPORTED_LANGUAGES.map((lang) => ({
+const LANGUAGE_OPTIONS = speech.SUPERTONIC_SUPPORTED_LANGUAGES.map((lang) => ({
   label: lang,
   value: lang,
 }));
@@ -60,7 +59,7 @@ function TTSContent() {
   const [text, setText] = useState(SAMPLE_TEXT);
   const [selectedModel, setSelectedModel] = useState<'XNNPACK_FP32' | 'MLX_FP32'>('XNNPACK_FP32');
   const [selectedVoice, setSelectedVoice] = useState<SupertonicDefaultVoiceName>('F1');
-  const [selectedLang, setSelectedLang] = useState<SupertonicLanguage>('en');
+  const [selectedLang, setSelectedLang] = useState<speech.SupertonicLanguage>('en');
   const [speed, setSpeed] = useState(1.05);
   const [totalSteps, setTotalSteps] = useState(8);
   const [isSynthesizing, setIsSynthesizing] = useState(false);
diff --git a/docs/typedoc.json b/docs/typedoc.json
index 30c55c65bf..b5552cf3ac 100644
--- a/docs/typedoc.json
+++ b/docs/typedoc.json
@@ -12,24 +12,17 @@
     "Core / Functions",
     "Core / Types",
     "Core / Errors",
-    "Core / Schema / Functions",
-    "Core / Schema / Types",
     "Utils / Functions",
     "Utils / Types",
     "CV / Tasks",
-    "CV / Functions",
     "CV / Types",
     "CV / Constants",
     "NLP / Tasks",
-    "NLP / Functions",
     "NLP / Types",
-    "NLP / Constants",
     "Speech / Tasks",
-    "Speech / Functions",
     "Speech / Types",
     "Speech / Constants",
     "LLM / Tasks",
-    "LLM / Functions",
     "LLM / Types",
     "*"
   ]
diff --git a/packages/react-native-executorch/package.json b/packages/react-native-executorch/package.json
index c15604ccea..a5b62933a2 100644
--- a/packages/react-native-executorch/package.json
+++ b/packages/react-native-executorch/package.json
@@ -13,6 +13,36 @@
       "types": "./lib/typescript/src/index.d.ts",
       "default": "./lib/module/index.js"
     },
+    "./cv": {
+      "source": "./src/extensions/cv/index.ts",
+      "types": "./lib/typescript/src/extensions/cv/index.d.ts",
+      "default": "./lib/module/extensions/cv/index.js"
+    },
+    "./llm": {
+      "source": "./src/extensions/llm/index.ts",
+      "types": "./lib/typescript/src/extensions/llm/index.d.ts",
+      "default": "./lib/module/extensions/llm/index.js"
+    },
+    "./nlp": {
+      "source": "./src/extensions/nlp/index.ts",
+      "types": "./lib/typescript/src/extensions/nlp/index.d.ts",
+      "default": "./lib/module/extensions/nlp/index.js"
+    },
+    "./speech": {
+      "source": "./src/extensions/speech/index.ts",
+      "types": "./lib/typescript/src/extensions/speech/index.d.ts",
+      "default": "./lib/module/extensions/speech/index.js"
+    },
+    "./math": {
+      "source": "./src/extensions/math.ts",
+      "types": "./lib/typescript/src/extensions/math.d.ts",
+      "default": "./lib/module/extensions/math.js"
+    },
+    "./schema": {
+      "source": "./src/core/schema.ts",
+      "types": "./lib/typescript/src/core/schema.d.ts",
+      "default": "./lib/module/core/schema.js"
+    },
     "./package.json": "./package.json"
   },
   "files": [
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index ac54e4c67c..1f6057a78a 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -10,7 +10,6 @@ import { scalePoint } from './points';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
- * @category CV / Types
  */
 export type BoxMap = Readonly<{
   xyxy: Readonly<{ xmin: number; ymin: number; xmax: number; ymax: number }>;
@@ -20,13 +19,11 @@ export type BoxMap = Readonly<{
 
 /**
  * The formats of bounding boxes.
- * @category CV / Types
  */
 export type BoxFormat = keyof BoxMap;
 
 /**
  * Representation of a bounding box under a specific format.
- * @category CV / Types
  */
 export type BoundingBox<F extends BoxFormat> = F extends any
   ? { readonly format: F } & BoxMap[F]
@@ -34,7 +31,6 @@ export type BoundingBox<F extends BoxFormat> = F extends any
 
 /**
  * Configuration options for scaling bounding box coordinates.
- * @category CV / Types
  */
 export type ScaleBoxOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -48,7 +44,6 @@ export type ScaleBoxOptions = {
 /**
  * Decodes bounding box coordinates from a 4-tuple into a structured BoundingBox
  * object.
- * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param tuple A 4-tuple array containing coordinates.
  * @param format The coordinate format to decode into.
@@ -72,7 +67,6 @@ export function decodeBox<F extends BoxFormat>(
 
 /**
  * Scales bounding box coordinates based on scaling options and resize modes.
- * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param box The original BoundingBox.
  * @param options Options defining dimensions and resize modes.
@@ -138,7 +132,6 @@ export function scaleBox<F extends BoxFormat>(
 
 /**
  * Options for Non-Maximum Suppression (NMS).
- * @category CV / Types
  */
 export type NmsOptions = {
   /** How bounding box coordinates are interpreted {@link BoxFormat}. */
@@ -157,7 +150,6 @@ export type NmsOptions = {
 /**
  * Executes Non-Maximum Suppression (NMS) on bounding boxes and confidence
  * scores.
- * @category CV / Functions
  * @param boxes Bounding boxes coordinate tensor. Expected shape `[N, 4]` and
  * data type `float32`.
  * @param scores Bounding boxes confidence scores tensor. Expected shape `[N]`
@@ -197,7 +189,6 @@ export function nms(boxes: Tensor, scores: Tensor, options: NmsOptions): number[
  * Note: This operation does not change the image tensor dimensions (it does not crop
  * the shape). Instead, it copies the elements within the box coordinates from
  * `src` to `dst`, and sets all elements outside the box to `0`.
- * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor to write masked values to.
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index e15421db1b..bd5bba4e2a 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -12,7 +12,6 @@ import type { ImageFormat } from '../image';
 
 /**
  * Supported color conversion code presets (similar to OpenCV).
- * @category CV / Types
  */
 export type ColorConversionCode =
   | 'RGBA2RGB'
@@ -67,19 +66,16 @@ export const FORMAT_CHANNELS: Record<ImageFormat, number> = {
 
 /**
  * Modes for resizing an image tensor to match target dimensions.
- * @category CV / Types
  */
 export type ResizeMode = 'stretch' | 'letterbox' | 'crop';
 
 /**
  * Interpolation algorithms used during image resizing.
- * @category CV / Types
  */
 export type InterpolationMethod = 'nearest' | 'area' | 'cubic' | 'lanczos' | 'linear';
 
 /**
  * Configuration options for image resize operations.
- * @category CV / Types
  */
 export type ResizeOptions = {
   /** How the image is resized (stretch, letterbox, or crop). */
@@ -92,7 +88,6 @@ export type ResizeOptions = {
 
 /**
  * Configuration options for image tensor normalization.
- * @category CV / Types
  */
 export type NormalizeOptions = {
   /**
@@ -140,7 +135,6 @@ export function resize(src: Tensor, dst: Tensor, options?: ResizeOptions): Tenso
 /**
  * Converts the color space of an image tensor using a specified color
  * conversion code.
- * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor to write the converted
@@ -165,7 +159,6 @@ export function cvtColor(src: Tensor, dst: Tensor, code: ColorConversionCode): T
  * (Channel, Height, Width).
  *
  * Commonly required for PyTorch Edge models which expect channels-first inputs.
- * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor in CHW layout. Expected
@@ -187,7 +180,6 @@ export function toChannelsFirst(src: Tensor, dst: Tensor): Tensor {
  *
  * Useful for post-processing model outputs back into channels-last layouts for
  * rendering or display.
- * @category CV / Functions
  * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
  * (channels-first). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor in HWC layout. Expected
@@ -209,7 +201,6 @@ export function toChannelsLast(src: Tensor, dst: Tensor): Tensor {
  * Computes: `dst[c,h,w] = src[c,h,w] * alpha[c] + beta[c]`. Can normalize
  * uniformly or channel-wise using array options. The result is cast to `dst`
  * tensor's dtype.
- * @category CV / Functions
  * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
  * (channels-first). Supports any numeric data type (typically `uint8` or
  * `float32`).
@@ -242,7 +233,6 @@ export function normalize(src: Tensor, dst: Tensor, options?: NormalizeOptions):
  * This operation iterates over each index/class ID in the source tensor, looks
  * up its corresponding RGBA color in the provided colormap palette, and writes
  * it to the destination tensor.
- * @category CV / Functions
  * @param src The source index/mask image tensor. Expected shape `[H, W, 1]` in
  * HWC layout with data type `int32` containing class indices.
  * @param dst The pre-allocated destination image tensor to write the mapped
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/index.ts b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
index e310574d16..4811cd2747 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
@@ -1,12 +1,7 @@
 /**
- * Aggregated image, box, and point operations.
+ * Aggregated image, box, point, and quad operations.
  */
 export * from './image';
 export * from './boxes';
 export * from './points';
 export * from './quad';
-
-export * as image from './image';
-export * as boxes from './boxes';
-export * as points from './points';
-export * as quad from './quad';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 54b0b2587b..a1f533b19a 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -6,7 +6,6 @@ import type { ResizeMode } from './image';
 
 /**
  * Represents a 2D coordinate point with x and y values.
- * @category CV / Types
  */
 export type Point = {
   readonly x: number;
@@ -15,7 +14,6 @@ export type Point = {
 
 /**
  * Euclidean distance between two points.
- * @category CV / Functions
  * @param a The first point.
  * @param b The second point.
  * @returns The distance between `a` and `b`.
@@ -28,7 +26,6 @@ export function distance(a: Point, b: Point): number {
 /**
  * Linearly interpolates between two points: `t = 0` returns `a`, `t = 1`
  * returns `b`, values in between interpolate along the segment.
- * @category CV / Functions
  * @param a The start point.
  * @param b The end point.
  * @param t The interpolation factor.
@@ -41,7 +38,6 @@ export function interpolatePoint(a: Point, b: Point, t: number): Point {
 
 /**
  * Configuration options for scaling 2D point coordinates.
- * @category CV / Types
  */
 export type ScalePointOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -55,7 +51,6 @@ export type ScalePointOptions = {
 /**
  * Helper function to scale a 2D point based on resize mode and resolution
  * changes.
- * @category CV / Functions
  * @param point The original coordinate point to scale.
  * @param options Options detailing the scaling factors and resize mode.
  * See {@link ScalePointOptions}.
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
index 69c8c88487..2404e69406 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -15,14 +15,12 @@ import type { ResizeMode } from './image';
  * Helpers that need them as top-left, top-right, bottom-right, bottom-left say
  * so on their `ordered` parameter; pass the quad through {@link orderQuad}
  * first. Never assume a `Quad` you were handed is already ordered.
- * @category CV / Types
  */
 export type Quad = readonly [Point, Point, Point, Point];
 
 /**
  * Computes the axis-aligned bounding box enclosing a set of points, in the
  * requested box format. Returns a zero box for empty input.
- * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param points The points to enclose.
  * @param format The coordinate format of the returned box.
@@ -73,7 +71,6 @@ export function boundingBoxOfPoints<F extends BoxFormat>(
  * Reorders a quad's corners into the top-left, top-right, bottom-right,
  * bottom-left order the rest of this module assumes, using their
  * coordinate-sum and coordinate-difference extremes.
- * @category CV / Functions
  * @param quad The quad whose corners may be in any order.
  * @returns The same corners, ordered TL, TR, BR, BL.
  */
@@ -98,7 +95,6 @@ export function orderQuad(quad: Quad): Quad {
 /**
  * Computes the width and height (in pixels) of an ordered TL,TR,BR,BL quad, taking
  * the longer of each pair of opposite sides.
- * @category CV / Functions
  * @param ordered The quad corners ordered TL, TR, BR, BL.
  * @returns The quad's width and height in pixels.
  */
@@ -112,7 +108,6 @@ export function quadSize(ordered: Quad): { width: number; height: number } {
 
 /**
  * Configuration options for scaling quad coordinates.
- * @category CV / Types
  */
 export type ScaleQuadOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -126,7 +121,6 @@ export type ScaleQuadOptions = {
 /**
  * Rescales a quad from one frame to another, clamping the result to the target
  * bounds. The counterpart of {@link scaleBox} for quads.
- * @category CV / Functions
  * @param quad The quad, expressed in the `from` frame.
  * @param options Options detailing the scaling factors and resize mode.
  * See {@link ScaleQuadOptions}.
@@ -144,7 +138,6 @@ export function scaleQuad(quad: Quad, options: ScaleQuadOptions): Quad {
 
 /**
  * Options for {@link rectifyQuad}.
- * @category CV / Types
  */
 export type RectifyQuadOptions = {
   /** Width in px the rectified content occupies inside the destination canvas. */
@@ -160,7 +153,6 @@ export type RectifyQuadOptions = {
  * `dst`: perspective crop, resize to the canvas height, and pad, in one native
  * pass. An axis-aligned bbox is a 4-corner quad, so pass its corners to
  * rectify a box.
- * @category CV / Functions
  * @param src The source image, `uint8` `[H, W, C]`.
  * @param dst The pre-allocated destination canvas, `uint8` `[H', W', C]`, with
  * the same channel count as `src`. Must not alias `src`.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index df3c4a63a3..e149d87c8c 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -24,8 +24,6 @@ import {
 } from '../ops/boxes';
 import { RnExecuTorchError } from '../../../core/error';
 
-export type { BoxFormat };
-
 /**
  * Options for configuring an instance segmenter preprocessor, label
  * vocabulary, and threshold parameters.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index 9cd2398d53..7bd42f2eb3 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -17,8 +17,6 @@ import type { ResizeMode } from '../ops/image';
 import { scalePoint, type Point } from '../ops/points';
 import { nms, type BoundingBox, type BoxFormat, decodeBox, scaleBox } from '../ops/boxes';
 
-export type { BoxFormat };
-
 /**
  * Options for configuring a keypoint detector runner.
  * @category CV / Types
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index 2c65bb5f63..0a7585aaf5 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -16,8 +16,6 @@ import { createImagePreprocessor, type ImagePreprocessorOptions } from './prepro
 import { nms, scaleBox, decodeBox, type BoundingBox, type BoxFormat } from '../ops/boxes';
 import { RnExecuTorchError } from '../../../core/error';
 
-export type { BoxFormat };
-
 /**
  * Options for configuring an object detector preprocessor, label vocabulary,
  * and detection thresholds.
diff --git a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
index 388cf065dd..89468ab202 100644
--- a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
@@ -30,7 +30,6 @@ function quadsFromFlat(flat: ArrayLike<number>): Quad[] {
 
 /**
  * Thresholds for {@link extractDbnetTextQuads}.
- * @category CV / Types
  */
 export type DbnetDecodeOptions = {
   /** Binarization threshold on the probability map. */
@@ -49,7 +48,6 @@ export type DbnetDecodeOptions = {
  * Decodes a DBNet probability map into oriented text quads: binarizes the map,
  * traces contours, scores each candidate by its mean probability and unclips the
  * survivors back to their unshrunk size.
- * @category CV / Functions
  * @param probabilityMap The `detect` output, shape `[1, 1, H, W]`, post-sigmoid.
  * @param options Decode thresholds. See {@link DbnetDecodeOptions}.
  * @returns The decoded quads, in detector-input pixel space and arbitrary order.
diff --git a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
index cfbb5a42d1..a75ee7704b 100644
--- a/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
+++ b/packages/react-native-executorch/src/extensions/llm/tasks/llmChatSession.ts
@@ -20,31 +20,11 @@ import {
 import { parseTokenizerConfig } from '../utils/tokenizerConfig';
 import {
   createChatPreprocessor,
-  type ChatMediaInput,
   type ChatMessageContent,
   type ChatMessage,
-  type LLMImagePreprocessorConfig,
-  type LLMAudioPreprocessorConfig,
   type LLMMediaPreprocessorConfig,
 } from '../utils/chatPreprocessor';
-import type { ToolDefinition, ToolCall, ToolParser, ToolParserResult } from '../utils/toolCalling';
-
-export type {
-  LLMKVCacheState,
-  LLMGenerationConfig,
-  LLMGenerationStats,
-  Modality,
-  ChatMediaInput,
-  ChatMessageContent,
-  ChatMessage,
-  LLMImagePreprocessorConfig,
-  LLMAudioPreprocessorConfig,
-  LLMMediaPreprocessorConfig,
-  ToolDefinition,
-  ToolCall,
-  ToolParser,
-  ToolParserResult,
-};
+import type { ToolDefinition, ToolParser } from '../utils/toolCalling';
 
 /**
  * Model configuration required to instantiate an LLM chat session.
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
index e22dd9861f..ed2482b2a3 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
@@ -17,11 +17,9 @@ import {
   buildGrammar,
   computeCharOffsets,
   extractSpans,
-  piiSegments,
   viterbiDecode,
   type PiiEntity,
   type PiiEntityType,
-  type PiiSegment,
   type ViterbiBiases,
 } from '../utils/privacyFilterUtils';
 
@@ -68,9 +66,6 @@ export type PrivacyFilterModel<Label extends string = string> = {
   readonly modelOpts: PrivacyFilterOptions<Label>;
 };
 
-export { piiSegments };
-export type { PiiEntity, PiiEntityType, PiiSegment, ViterbiBiases };
-
 /**
  * Privacy filter task runner for detecting PII entities in text.
  * @category NLP / Types
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
index 62129979ee..903b5013f7 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
@@ -17,15 +17,12 @@ import {
   formatChunk,
   encodeText,
   parseVoiceStyle,
-  SUPERTONIC_SUPPORTED_LANGUAGES,
   type SupertonicVoiceStyle,
   type SupertonicLanguage,
 } from '../utils/supertonicUtils';
 import { partition } from '../utils/textPartitioner';
 import { RnExecuTorchError } from '../../../core/error';
 
-export { SUPERTONIC_SUPPORTED_LANGUAGES, type SupertonicVoiceStyle, type SupertonicLanguage };
-
 /**
  * Supertonic audio sampling rate in Hz (44100).
  * @category Speech / Constants
diff --git a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
index e91c015767..1698916e48 100644
--- a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
@@ -3,8 +3,8 @@ import { useResourceDownload, type ResourceOptions } from './useResourceDownload
 import {
   createInstanceSegmenter,
   type InstanceSegmenterModel,
-  type BoxFormat,
 } from '../extensions/cv/tasks/instanceSegmentation';
+import type { BoxFormat } from '../extensions/cv/ops/boxes';
 
 /**
  * React hook to load and run an instance segmentation model.
diff --git a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
index 89ec0012f8..20cc6d20e5 100644
--- a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
@@ -3,8 +3,8 @@ import { useResourceDownload, type ResourceOptions } from './useResourceDownload
 import {
   createKeypointDetector,
   type KeypointDetectorModel,
-  type BoxFormat,
 } from '../extensions/cv/tasks/keypointDetection';
+import type { BoxFormat } from '../extensions/cv/ops/boxes';
 
 /**
  * React hook to load and run a keypoint detection model.
diff --git a/packages/react-native-executorch/src/hooks/useObjectDetector.ts b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
index 4acbccd4da..bc1301e3ed 100644
--- a/packages/react-native-executorch/src/hooks/useObjectDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
@@ -3,8 +3,8 @@ import { useResourceDownload, type ResourceOptions } from './useResourceDownload
 import {
   createObjectDetector,
   type ObjectDetectorModel,
-  type BoxFormat,
 } from '../extensions/cv/tasks/objectDetection';
+import type { BoxFormat } from '../extensions/cv/ops/boxes';
 
 /**
  * React hook to load and run an object detection model.
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 2b8ba5b91b..b9a8c4a278 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -98,8 +98,6 @@ export * from './core/error';
 export * from './core/model';
 export * from './core/tensor';
 export * from './core/runtime';
-
-export type * from './core/schema';
 /**
  * Model schema validation and dimension constraints.
  * @category Modules

From b8b3ef80e41eab7ed776b3eb9f5ce0e0351ef452 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:10:26 +0200
Subject: [PATCH 23/38] chore: move preprocessing to
 cv/utils/imagePreprocessor, remove worklet from dispose

---
 docs/typedoc.json                                      |  2 +-
 packages/react-native-executorch/src/core/error.ts     | 10 +++++-----
 .../react-native-executorch/src/extensions/cv/index.ts |  2 +-
 .../src/extensions/cv/ops/image.ts                     |  1 -
 .../src/extensions/cv/tasks/classification.ts          |  2 +-
 .../src/extensions/cv/tasks/imageEmbedding.ts          |  2 +-
 .../src/extensions/cv/tasks/instanceSegmentation.ts    |  2 +-
 .../src/extensions/cv/tasks/keypointDetection.ts       |  2 +-
 .../src/extensions/cv/tasks/objectDetection.ts         |  2 +-
 .../src/extensions/cv/tasks/paddleOcr.ts               |  2 +-
 .../src/extensions/cv/tasks/semanticSegmentation.ts    |  2 +-
 .../src/extensions/cv/tasks/styleTransfer.ts           |  2 +-
 .../preprocessing.ts => utils/imagePreprocessor.ts}    |  4 ----
 .../src/extensions/llm/utils/chatPreprocessor.ts       |  2 +-
 packages/react-native-executorch/src/index.ts          |  1 -
 15 files changed, 16 insertions(+), 22 deletions(-)
 rename packages/react-native-executorch/src/extensions/cv/{tasks/preprocessing.ts => utils/imagePreprocessor.ts} (98%)

diff --git a/docs/typedoc.json b/docs/typedoc.json
index b5552cf3ac..c0339a4683 100644
--- a/docs/typedoc.json
+++ b/docs/typedoc.json
@@ -11,7 +11,6 @@
     "Modules",
     "Core / Functions",
     "Core / Types",
-    "Core / Errors",
     "Utils / Functions",
     "Utils / Types",
     "CV / Tasks",
@@ -19,6 +18,7 @@
     "CV / Constants",
     "NLP / Tasks",
     "NLP / Types",
+    "NLP / Constants",
     "Speech / Tasks",
     "Speech / Types",
     "Speech / Constants",
diff --git a/packages/react-native-executorch/src/core/error.ts b/packages/react-native-executorch/src/core/error.ts
index a3d2a529b3..84374a6967 100644
--- a/packages/react-native-executorch/src/core/error.ts
+++ b/packages/react-native-executorch/src/core/error.ts
@@ -18,7 +18,7 @@
  * resource, re-create a disposed one). Everything else is a category that
  * exists so crash reporters can group failures, and the detail lives in the
  * message.
- * @category Core / Errors
+ * @category Core / Types
  */
 export const VALID_ERROR_CODES = [
   'LOAD_FAILED',
@@ -37,7 +37,7 @@ export const VALID_ERROR_CODES = [
  * Machine-readable classification of an {@link RnExecuTorchError}. Branch on
  * this rather than on the message, which is written for humans and can be
  * reworded in any release.
- * @category Core / Errors
+ * @category Core / Types
  */
 export type RnExecuTorchErrorCode = (typeof VALID_ERROR_CODES)[number];
 
@@ -53,7 +53,7 @@ export type RnExecuTorchErrorCode = (typeof VALID_ERROR_CODES)[number];
  * ```typescript
  * throw RnExecuTorchError('INVALID_ARGUMENT', 'Shape dimensions must be positive');
  * ```
- * @category Core / Errors
+ * @category Core / Types
  * @typeParam C The specific code, narrowed by {@link isRnExecuTorchError}.
  */
 export type RnExecuTorchError<C extends RnExecuTorchErrorCode = RnExecuTorchErrorCode> = Error & {
@@ -69,7 +69,7 @@ export type RnExecuTorchError<C extends RnExecuTorchErrorCode = RnExecuTorchErro
 
 /**
  * Creates an {@link RnExecuTorchError} carrying a machine-readable code.
- * @category Core / Errors
+ * @category Core / Functions
  * @typeParam C The specific error code.
  * @param code The classification to attach. See {@link RnExecuTorchErrorCode}.
  * @param message A human-readable description of the failure.
@@ -104,7 +104,7 @@ export function RnExecuTorchError<C extends RnExecuTorchErrorCode>(
  *
  * Duck-typed so it holds for errors that crossed a worklet or JSI boundary,
  * where class identity is gone.
- * @category Core / Errors
+ * @category Core / Functions
  * @param err The caught value.
  * @param code When given, also requires the error to carry exactly this code.
  * @returns Whether `err` is an `RnExecuTorchError` (of code `code`, if given).
diff --git a/packages/react-native-executorch/src/extensions/cv/index.ts b/packages/react-native-executorch/src/extensions/cv/index.ts
index a3d21e4a06..6370cfd125 100644
--- a/packages/react-native-executorch/src/extensions/cv/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/index.ts
@@ -6,4 +6,4 @@
 export * from './image';
 export * from './ops';
 export * from './utils/paddleOcrUtils';
-export * from './tasks/preprocessing';
+export * from './utils/imagePreprocessor';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index bd5bba4e2a..fec1e365c2 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -106,7 +106,6 @@ export type NormalizeOptions = {
  * Resizes an image tensor from a source dimension to a destination dimension.
  *
  * Supports various {@link ResizeMode} and {@link InterpolationMethod} options.
- * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type (e.g. `uint8`, `float32`).
  * @param dst The pre-allocated destination image tensor to write the resized
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
index efbc2cf36b..bb4c059c3e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/classification.ts
@@ -12,7 +12,7 @@ import { wrapAsync } from '../../../core/runtime';
 
 import { softmax } from '../../math';
 import type { ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 import { RnExecuTorchError } from '../../../core/error';
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
index a2e95f07ff..868e3bdada 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/imageEmbedding.ts
@@ -10,7 +10,7 @@ import { validateSpec, method, f32 } from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 
 import type { ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 
 /**
  * Model configuration required to instantiate an image embedder task runner.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index e149d87c8c..b1f8cc0160 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -11,7 +11,7 @@ import { validateSpec, method, f32 } from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 
 import type { ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 import { threshold } from '../../math';
 import { resize, normalize } from '../ops/image';
 import {
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index 7bd42f2eb3..e3dabc5723 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -11,7 +11,7 @@ import { validateSpec, method, f32 } from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 
 import type { ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 
 import type { ResizeMode } from '../ops/image';
 import { scalePoint, type Point } from '../ops/points';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index 0a7585aaf5..a9ac645037 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -12,7 +12,7 @@ import { wrapAsync } from '../../../core/runtime';
 
 import type { ResizeMode } from '../ops/image';
 import type { ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 import { nms, scaleBox, decodeBox, type BoundingBox, type BoxFormat } from '../ops/boxes';
 import { RnExecuTorchError } from '../../../core/error';
 
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 7e1c4f0ec3..28bd9d7932 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -39,7 +39,7 @@ import {
 } from '../ops/quad';
 import { argmax, gather } from '../../math';
 import { extractDbnetTextQuads } from '../utils/paddleOcrUtils';
-import { createImagePreprocessor } from './preprocessing';
+import { createImagePreprocessor } from '../utils/imagePreprocessor';
 
 /**
  * A single recognized text region.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
index dea0ec5c0e..7cf25928b6 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/semanticSegmentation.ts
@@ -11,7 +11,7 @@ import { validateSpec, method, f32 } from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 
 import type { ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 import {
   toChannelsLast,
   normalize,
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
index 8a537ad87a..2232c0eff6 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/styleTransfer.ts
@@ -11,7 +11,7 @@ import { validateSpec, method, f32 } from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 
 import type { ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 import {
   toChannelsLast,
   normalize,
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts b/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
similarity index 98%
rename from packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
rename to packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
index 6970047fd9..c59c6f619f 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
@@ -20,7 +20,6 @@ import { RnExecuTorchError } from '../../../core/error';
 
 /**
  * Options for configuring the image preprocessor pipeline.
- * @category CV / Types
  */
 export type ImagePreprocessorOptions = {
   /** How the input image is resized to match the model's expected dimensions. */
@@ -35,7 +34,6 @@ export type ImagePreprocessorOptions = {
 
 /**
  * Image preprocessor runner for transforming image buffers into model input tensors.
- * @category CV / Types
  */
 export type ImagePreprocessor = {
   /**
@@ -64,7 +62,6 @@ export type ImagePreprocessor = {
  * normalize, and copy raw image buffers into target tensors matching model
  * input shapes. All intermediate scratch tensors are pre-allocated and safely
  * disposed of when calling `dispose()`.
- * @category CV / Functions
  * @param options Normalization scaling coefficients, interpolation algorithms, and
  * resize modes.
  * See {@link ImagePreprocessorOptions}.
@@ -102,7 +99,6 @@ export function createImagePreprocessor(
   const { resizeMode, interpolation, normalizeOpts, padValue } = options;
 
   const dispose = () => {
-    'worklet';
     tensors.forEach((t) => t.dispose());
   };
   const process = (input: ImageBuffer): Tensor => {
diff --git a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
index 8389657f8e..77b5b37ab0 100644
--- a/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
+++ b/packages/react-native-executorch/src/extensions/llm/utils/chatPreprocessor.ts
@@ -11,7 +11,7 @@ import type { ImageBuffer } from '../../cv';
 import {
   createImagePreprocessor,
   type ImagePreprocessorOptions,
-} from '../../cv/tasks/preprocessing';
+} from '../../cv/utils/imagePreprocessor';
 
 import type { Modality, Prompt, MediaInput } from '../llmRunner';
 import type { ToolDefinition, ToolCall } from './toolCalling';
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index b9a8c4a278..d6fccf27ed 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -83,7 +83,6 @@ export * from './extensions/cv/tasks/objectDetection';
 export * from './extensions/cv/tasks/imageEmbedding';
 export * from './extensions/cv/tasks/sdxsTextToImage';
 export * from './extensions/cv/tasks/paddleOcr';
-export * from './extensions/cv/tasks/preprocessing';
 export * from './extensions/llm/tasks/llmChatSession';
 export * from './extensions/nlp/tasks/tokenization';
 export * from './extensions/nlp/tasks/textEmbedding';

From 7d21bf890723f95358949820a89b343640db213e Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:11:15 +0200
Subject: [PATCH 24/38] style

---
 .../src/extensions/cv/utils/imagePreprocessor.ts             | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts b/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
index c59c6f619f..45dc5042d4 100644
--- a/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
@@ -98,9 +98,8 @@ export function createImagePreprocessor(
   const [tColor, tChanFirst, tNorm, tOutput] = tensors;
   const { resizeMode, interpolation, normalizeOpts, padValue } = options;
 
-  const dispose = () => {
-    tensors.forEach((t) => t.dispose());
-  };
+  const dispose = () => tensors.forEach((t) => t.dispose());
+
   const process = (input: ImageBuffer): Tensor => {
     'worklet';
     const { data, width, height, format } = input;

From 33d394921fe2ee843acf0e90aa28a02ee6962ca8 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:12:51 +0200
Subject: [PATCH 25/38] chore: rename ops/boxes -> box, ops/points -> point
 (singular consistency)

---
 .../src/extensions/cv/ops/{boxes.ts => box.ts}                | 2 +-
 .../react-native-executorch/src/extensions/cv/ops/index.ts    | 4 ++--
 .../src/extensions/cv/ops/{points.ts => point.ts}             | 0
 .../react-native-executorch/src/extensions/cv/ops/quad.ts     | 4 ++--
 .../src/extensions/cv/tasks/instanceSegmentation.ts           | 2 +-
 .../src/extensions/cv/tasks/keypointDetection.ts              | 4 ++--
 .../src/extensions/cv/tasks/objectDetection.ts                | 2 +-
 .../src/extensions/cv/tasks/paddleOcr.ts                      | 2 +-
 .../react-native-executorch/src/hooks/useInstanceSegmenter.ts | 2 +-
 .../react-native-executorch/src/hooks/useKeypointDetector.ts  | 2 +-
 .../react-native-executorch/src/hooks/useObjectDetector.ts    | 2 +-
 11 files changed, 13 insertions(+), 13 deletions(-)
 rename packages/react-native-executorch/src/extensions/cv/ops/{boxes.ts => box.ts} (99%)
 rename packages/react-native-executorch/src/extensions/cv/ops/{points.ts => point.ts} (100%)

diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/box.ts
similarity index 99%
rename from packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
rename to packages/react-native-executorch/src/extensions/cv/ops/box.ts
index 1f6057a78a..75d811c55e 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/box.ts
@@ -6,7 +6,7 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ResizeMode } from './image';
-import { scalePoint } from './points';
+import { scalePoint } from './point';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/index.ts b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
index 4811cd2747..f3aa2c8ecf 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
@@ -2,6 +2,6 @@
  * Aggregated image, box, point, and quad operations.
  */
 export * from './image';
-export * from './boxes';
-export * from './points';
+export * from './box';
+export * from './point';
 export * from './quad';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/point.ts
similarity index 100%
rename from packages/react-native-executorch/src/extensions/cv/ops/points.ts
rename to packages/react-native-executorch/src/extensions/cv/ops/point.ts
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
index 2404e69406..31213da860 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -6,8 +6,8 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import { RnExecuTorchError } from '../../../core/error';
-import { distance, scalePoint, type Point } from './points';
-import type { BoundingBox, BoxFormat } from './boxes';
+import { distance, scalePoint, type Point } from './point';
+import type { BoundingBox, BoxFormat } from './box';
 import type { ResizeMode } from './image';
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index b1f8cc0160..d4bc55fe0a 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -21,7 +21,7 @@ import {
   restrictToBox,
   type BoundingBox,
   type BoxFormat,
-} from '../ops/boxes';
+} from '../ops/box';
 import { RnExecuTorchError } from '../../../core/error';
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index e3dabc5723..2cd7771610 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -14,8 +14,8 @@ import type { ImageBuffer } from '../image';
 import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 
 import type { ResizeMode } from '../ops/image';
-import { scalePoint, type Point } from '../ops/points';
-import { nms, type BoundingBox, type BoxFormat, decodeBox, scaleBox } from '../ops/boxes';
+import { scalePoint, type Point } from '../ops/point';
+import { nms, type BoundingBox, type BoxFormat, decodeBox, scaleBox } from '../ops/box';
 
 /**
  * Options for configuring a keypoint detector runner.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index a9ac645037..7760ce05fa 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -13,7 +13,7 @@ import { wrapAsync } from '../../../core/runtime';
 import type { ResizeMode } from '../ops/image';
 import type { ImageBuffer } from '../image';
 import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
-import { nms, scaleBox, decodeBox, type BoundingBox, type BoxFormat } from '../ops/boxes';
+import { nms, scaleBox, decodeBox, type BoundingBox, type BoxFormat } from '../ops/box';
 import { RnExecuTorchError } from '../../../core/error';
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 28bd9d7932..5472e559e5 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -28,7 +28,7 @@ import {
   FORMAT_CHANNELS,
   FORMAT_CONVERSION,
 } from '../ops/image';
-import { interpolatePoint } from '../ops/points';
+import { interpolatePoint } from '../ops/point';
 import {
   boundingBoxOfPoints,
   orderQuad,
diff --git a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
index 1698916e48..06f292c750 100644
--- a/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
+++ b/packages/react-native-executorch/src/hooks/useInstanceSegmenter.ts
@@ -4,7 +4,7 @@ import {
   createInstanceSegmenter,
   type InstanceSegmenterModel,
 } from '../extensions/cv/tasks/instanceSegmentation';
-import type { BoxFormat } from '../extensions/cv/ops/boxes';
+import type { BoxFormat } from '../extensions/cv/ops/box';
 
 /**
  * React hook to load and run an instance segmentation model.
diff --git a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
index 20cc6d20e5..ab3147fd58 100644
--- a/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useKeypointDetector.ts
@@ -4,7 +4,7 @@ import {
   createKeypointDetector,
   type KeypointDetectorModel,
 } from '../extensions/cv/tasks/keypointDetection';
-import type { BoxFormat } from '../extensions/cv/ops/boxes';
+import type { BoxFormat } from '../extensions/cv/ops/box';
 
 /**
  * React hook to load and run a keypoint detection model.
diff --git a/packages/react-native-executorch/src/hooks/useObjectDetector.ts b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
index bc1301e3ed..eea3889f9d 100644
--- a/packages/react-native-executorch/src/hooks/useObjectDetector.ts
+++ b/packages/react-native-executorch/src/hooks/useObjectDetector.ts
@@ -4,7 +4,7 @@ import {
   createObjectDetector,
   type ObjectDetectorModel,
 } from '../extensions/cv/tasks/objectDetection';
-import type { BoxFormat } from '../extensions/cv/ops/boxes';
+import type { BoxFormat } from '../extensions/cv/ops/box';
 
 /**
  * React hook to load and run an object detection model.

From daddb4e4644b126fc65c9432a7eb5a965c35e5a8 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:19:32 +0200
Subject: [PATCH 26/38] fix: add subpath tsconfig path aliases for all app
 tsconfigs

---
 apps/computer-vision/tsconfig.json | 12 +++++++++++-
 apps/nlp/tsconfig.json             | 12 +++++++++++-
 apps/speech/tsconfig.json          | 12 +++++++++++-
 3 files changed, 33 insertions(+), 3 deletions(-)

diff --git a/apps/computer-vision/tsconfig.json b/apps/computer-vision/tsconfig.json
index 47026ce434..9baba0ef64 100644
--- a/apps/computer-vision/tsconfig.json
+++ b/apps/computer-vision/tsconfig.json
@@ -9,7 +9,17 @@
     "customConditions": ["react-native"],
     "noEmit": true,
     "paths": {
-      "react-native-executorch": ["../../packages/react-native-executorch/src"]
+      "react-native-executorch": ["../../packages/react-native-executorch/src"],
+      "react-native-executorch/cv": ["../../packages/react-native-executorch/src/extensions/cv"],
+      "react-native-executorch/llm": ["../../packages/react-native-executorch/src/extensions/llm"],
+      "react-native-executorch/nlp": ["../../packages/react-native-executorch/src/extensions/nlp"],
+      "react-native-executorch/speech": [
+        "../../packages/react-native-executorch/src/extensions/speech"
+      ],
+      "react-native-executorch/math": [
+        "../../packages/react-native-executorch/src/extensions/math"
+      ],
+      "react-native-executorch/schema": ["../../packages/react-native-executorch/src/core/schema"]
     }
   }
 }
diff --git a/apps/nlp/tsconfig.json b/apps/nlp/tsconfig.json
index 47026ce434..9baba0ef64 100644
--- a/apps/nlp/tsconfig.json
+++ b/apps/nlp/tsconfig.json
@@ -9,7 +9,17 @@
     "customConditions": ["react-native"],
     "noEmit": true,
     "paths": {
-      "react-native-executorch": ["../../packages/react-native-executorch/src"]
+      "react-native-executorch": ["../../packages/react-native-executorch/src"],
+      "react-native-executorch/cv": ["../../packages/react-native-executorch/src/extensions/cv"],
+      "react-native-executorch/llm": ["../../packages/react-native-executorch/src/extensions/llm"],
+      "react-native-executorch/nlp": ["../../packages/react-native-executorch/src/extensions/nlp"],
+      "react-native-executorch/speech": [
+        "../../packages/react-native-executorch/src/extensions/speech"
+      ],
+      "react-native-executorch/math": [
+        "../../packages/react-native-executorch/src/extensions/math"
+      ],
+      "react-native-executorch/schema": ["../../packages/react-native-executorch/src/core/schema"]
     }
   }
 }
diff --git a/apps/speech/tsconfig.json b/apps/speech/tsconfig.json
index 47026ce434..9baba0ef64 100644
--- a/apps/speech/tsconfig.json
+++ b/apps/speech/tsconfig.json
@@ -9,7 +9,17 @@
     "customConditions": ["react-native"],
     "noEmit": true,
     "paths": {
-      "react-native-executorch": ["../../packages/react-native-executorch/src"]
+      "react-native-executorch": ["../../packages/react-native-executorch/src"],
+      "react-native-executorch/cv": ["../../packages/react-native-executorch/src/extensions/cv"],
+      "react-native-executorch/llm": ["../../packages/react-native-executorch/src/extensions/llm"],
+      "react-native-executorch/nlp": ["../../packages/react-native-executorch/src/extensions/nlp"],
+      "react-native-executorch/speech": [
+        "../../packages/react-native-executorch/src/extensions/speech"
+      ],
+      "react-native-executorch/math": [
+        "../../packages/react-native-executorch/src/extensions/math"
+      ],
+      "react-native-executorch/schema": ["../../packages/react-native-executorch/src/core/schema"]
     }
   }
 }

From 40be915e17f1e0863af155d5769e1bf3247c5b3d Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:29:22 +0200
Subject: [PATCH 27/38] docs: add @category tags to all CV ops and missing NLP
 utils functions

---
 .../src/extensions/cv/ops/box.ts                      |  9 +++++++++
 .../src/extensions/cv/ops/image.ts                    | 11 +++++++++++
 .../src/extensions/cv/ops/point.ts                    |  5 +++++
 .../src/extensions/cv/ops/quad.ts                     |  8 ++++++++
 .../src/extensions/nlp/utils/privacyFilterUtils.ts    |  5 +++++
 5 files changed, 38 insertions(+)

diff --git a/packages/react-native-executorch/src/extensions/cv/ops/box.ts b/packages/react-native-executorch/src/extensions/cv/ops/box.ts
index 75d811c55e..61da7dae1e 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/box.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/box.ts
@@ -10,6 +10,7 @@ import { scalePoint } from './point';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
+ * @category CV / Types
  */
 export type BoxMap = Readonly<{
   xyxy: Readonly<{ xmin: number; ymin: number; xmax: number; ymax: number }>;
@@ -19,11 +20,13 @@ export type BoxMap = Readonly<{
 
 /**
  * The formats of bounding boxes.
+ * @category CV / Types
  */
 export type BoxFormat = keyof BoxMap;
 
 /**
  * Representation of a bounding box under a specific format.
+ * @category CV / Types
  */
 export type BoundingBox<F extends BoxFormat> = F extends any
   ? { readonly format: F } & BoxMap[F]
@@ -31,6 +34,7 @@ export type BoundingBox<F extends BoxFormat> = F extends any
 
 /**
  * Configuration options for scaling bounding box coordinates.
+ * @category CV / Types
  */
 export type ScaleBoxOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -44,6 +48,7 @@ export type ScaleBoxOptions = {
 /**
  * Decodes bounding box coordinates from a 4-tuple into a structured BoundingBox
  * object.
+ * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param tuple A 4-tuple array containing coordinates.
  * @param format The coordinate format to decode into.
@@ -67,6 +72,7 @@ export function decodeBox<F extends BoxFormat>(
 
 /**
  * Scales bounding box coordinates based on scaling options and resize modes.
+ * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param box The original BoundingBox.
  * @param options Options defining dimensions and resize modes.
@@ -132,6 +138,7 @@ export function scaleBox<F extends BoxFormat>(
 
 /**
  * Options for Non-Maximum Suppression (NMS).
+ * @category CV / Types
  */
 export type NmsOptions = {
   /** How bounding box coordinates are interpreted {@link BoxFormat}. */
@@ -150,6 +157,7 @@ export type NmsOptions = {
 /**
  * Executes Non-Maximum Suppression (NMS) on bounding boxes and confidence
  * scores.
+ * @category CV / Functions
  * @param boxes Bounding boxes coordinate tensor. Expected shape `[N, 4]` and
  * data type `float32`.
  * @param scores Bounding boxes confidence scores tensor. Expected shape `[N]`
@@ -189,6 +197,7 @@ export function nms(boxes: Tensor, scores: Tensor, options: NmsOptions): number[
  * Note: This operation does not change the image tensor dimensions (it does not crop
  * the shape). Instead, it copies the elements within the box coordinates from
  * `src` to `dst`, and sets all elements outside the box to `0`.
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor to write masked values to.
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index fec1e365c2..e15421db1b 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -12,6 +12,7 @@ import type { ImageFormat } from '../image';
 
 /**
  * Supported color conversion code presets (similar to OpenCV).
+ * @category CV / Types
  */
 export type ColorConversionCode =
   | 'RGBA2RGB'
@@ -66,16 +67,19 @@ export const FORMAT_CHANNELS: Record<ImageFormat, number> = {
 
 /**
  * Modes for resizing an image tensor to match target dimensions.
+ * @category CV / Types
  */
 export type ResizeMode = 'stretch' | 'letterbox' | 'crop';
 
 /**
  * Interpolation algorithms used during image resizing.
+ * @category CV / Types
  */
 export type InterpolationMethod = 'nearest' | 'area' | 'cubic' | 'lanczos' | 'linear';
 
 /**
  * Configuration options for image resize operations.
+ * @category CV / Types
  */
 export type ResizeOptions = {
   /** How the image is resized (stretch, letterbox, or crop). */
@@ -88,6 +92,7 @@ export type ResizeOptions = {
 
 /**
  * Configuration options for image tensor normalization.
+ * @category CV / Types
  */
 export type NormalizeOptions = {
   /**
@@ -106,6 +111,7 @@ export type NormalizeOptions = {
  * Resizes an image tensor from a source dimension to a destination dimension.
  *
  * Supports various {@link ResizeMode} and {@link InterpolationMethod} options.
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type (e.g. `uint8`, `float32`).
  * @param dst The pre-allocated destination image tensor to write the resized
@@ -134,6 +140,7 @@ export function resize(src: Tensor, dst: Tensor, options?: ResizeOptions): Tenso
 /**
  * Converts the color space of an image tensor using a specified color
  * conversion code.
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor to write the converted
@@ -158,6 +165,7 @@ export function cvtColor(src: Tensor, dst: Tensor, code: ColorConversionCode): T
  * (Channel, Height, Width).
  *
  * Commonly required for PyTorch Edge models which expect channels-first inputs.
+ * @category CV / Functions
  * @param src The source image tensor in HWC layout. Expected shape `[H, W, C]`
  * (channels-last). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor in CHW layout. Expected
@@ -179,6 +187,7 @@ export function toChannelsFirst(src: Tensor, dst: Tensor): Tensor {
  *
  * Useful for post-processing model outputs back into channels-last layouts for
  * rendering or display.
+ * @category CV / Functions
  * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
  * (channels-first). Supports any numeric data type.
  * @param dst The pre-allocated destination image tensor in HWC layout. Expected
@@ -200,6 +209,7 @@ export function toChannelsLast(src: Tensor, dst: Tensor): Tensor {
  * Computes: `dst[c,h,w] = src[c,h,w] * alpha[c] + beta[c]`. Can normalize
  * uniformly or channel-wise using array options. The result is cast to `dst`
  * tensor's dtype.
+ * @category CV / Functions
  * @param src The source image tensor in CHW layout. Expected shape `[C, H, W]`
  * (channels-first). Supports any numeric data type (typically `uint8` or
  * `float32`).
@@ -232,6 +242,7 @@ export function normalize(src: Tensor, dst: Tensor, options?: NormalizeOptions):
  * This operation iterates over each index/class ID in the source tensor, looks
  * up its corresponding RGBA color in the provided colormap palette, and writes
  * it to the destination tensor.
+ * @category CV / Functions
  * @param src The source index/mask image tensor. Expected shape `[H, W, 1]` in
  * HWC layout with data type `int32` containing class indices.
  * @param dst The pre-allocated destination image tensor to write the mapped
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/point.ts b/packages/react-native-executorch/src/extensions/cv/ops/point.ts
index a1f533b19a..54b0b2587b 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/point.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/point.ts
@@ -6,6 +6,7 @@ import type { ResizeMode } from './image';
 
 /**
  * Represents a 2D coordinate point with x and y values.
+ * @category CV / Types
  */
 export type Point = {
   readonly x: number;
@@ -14,6 +15,7 @@ export type Point = {
 
 /**
  * Euclidean distance between two points.
+ * @category CV / Functions
  * @param a The first point.
  * @param b The second point.
  * @returns The distance between `a` and `b`.
@@ -26,6 +28,7 @@ export function distance(a: Point, b: Point): number {
 /**
  * Linearly interpolates between two points: `t = 0` returns `a`, `t = 1`
  * returns `b`, values in between interpolate along the segment.
+ * @category CV / Functions
  * @param a The start point.
  * @param b The end point.
  * @param t The interpolation factor.
@@ -38,6 +41,7 @@ export function interpolatePoint(a: Point, b: Point, t: number): Point {
 
 /**
  * Configuration options for scaling 2D point coordinates.
+ * @category CV / Types
  */
 export type ScalePointOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -51,6 +55,7 @@ export type ScalePointOptions = {
 /**
  * Helper function to scale a 2D point based on resize mode and resolution
  * changes.
+ * @category CV / Functions
  * @param point The original coordinate point to scale.
  * @param options Options detailing the scaling factors and resize mode.
  * See {@link ScalePointOptions}.
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
index 31213da860..79b705aced 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -15,12 +15,14 @@ import type { ResizeMode } from './image';
  * Helpers that need them as top-left, top-right, bottom-right, bottom-left say
  * so on their `ordered` parameter; pass the quad through {@link orderQuad}
  * first. Never assume a `Quad` you were handed is already ordered.
+ * @category CV / Types
  */
 export type Quad = readonly [Point, Point, Point, Point];
 
 /**
  * Computes the axis-aligned bounding box enclosing a set of points, in the
  * requested box format. Returns a zero box for empty input.
+ * @category CV / Functions
  * @typeParam F Bounding box coordinate format.
  * @param points The points to enclose.
  * @param format The coordinate format of the returned box.
@@ -71,6 +73,7 @@ export function boundingBoxOfPoints<F extends BoxFormat>(
  * Reorders a quad's corners into the top-left, top-right, bottom-right,
  * bottom-left order the rest of this module assumes, using their
  * coordinate-sum and coordinate-difference extremes.
+ * @category CV / Functions
  * @param quad The quad whose corners may be in any order.
  * @returns The same corners, ordered TL, TR, BR, BL.
  */
@@ -95,6 +98,7 @@ export function orderQuad(quad: Quad): Quad {
 /**
  * Computes the width and height (in pixels) of an ordered TL,TR,BR,BL quad, taking
  * the longer of each pair of opposite sides.
+ * @category CV / Functions
  * @param ordered The quad corners ordered TL, TR, BR, BL.
  * @returns The quad's width and height in pixels.
  */
@@ -108,6 +112,7 @@ export function quadSize(ordered: Quad): { width: number; height: number } {
 
 /**
  * Configuration options for scaling quad coordinates.
+ * @category CV / Types
  */
 export type ScaleQuadOptions = {
   /** The source bounds (e.g. model input dimensions). */
@@ -121,6 +126,7 @@ export type ScaleQuadOptions = {
 /**
  * Rescales a quad from one frame to another, clamping the result to the target
  * bounds. The counterpart of {@link scaleBox} for quads.
+ * @category CV / Functions
  * @param quad The quad, expressed in the `from` frame.
  * @param options Options detailing the scaling factors and resize mode.
  * See {@link ScaleQuadOptions}.
@@ -138,6 +144,7 @@ export function scaleQuad(quad: Quad, options: ScaleQuadOptions): Quad {
 
 /**
  * Options for {@link rectifyQuad}.
+ * @category CV / Types
  */
 export type RectifyQuadOptions = {
   /** Width in px the rectified content occupies inside the destination canvas. */
@@ -153,6 +160,7 @@ export type RectifyQuadOptions = {
  * `dst`: perspective crop, resize to the canvas height, and pad, in one native
  * pass. An axis-aligned bbox is a 4-corner quad, so pass its corners to
  * rectify a box.
+ * @category CV / Functions
  * @param src The source image, `uint8` `[H, W, C]`.
  * @param dst The pre-allocated destination canvas, `uint8` `[H', W', C]`, with
  * the same channel count as `src`. Must not alias `src`.
diff --git a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
index e4e20601d6..c0384f081d 100644
--- a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
@@ -138,6 +138,7 @@ function classOf(name: string): number {
  * The encoded grammar is:
  * - `O` / `E-x` / `S-x` -> `O` | `B-y` | `S-y`
  * - `B-x` / `I-x` -> `I-x` | `E-x` (same entity)
+ * @category NLP / Functions
  * @param labelNames BIOES label list; index 0 must be `'O'`.
  * @param biases Optional transition biases; missing fields default to `0`.
  * @returns The pre-computed grammar.
@@ -208,6 +209,7 @@ export function buildGrammar(labelNames: readonly string[], biases?: ViterbiBias
  * Runs constrained Viterbi over a `[validLen, numLabels]` slice of per-token
  * logits and returns the best BIOES-grammar-valid label-id sequence (length
  * `validLen`).
+ * @category NLP / Functions
  * @param logits Flat row-major logits; row `t` starts at `t * numLabels`.
  * @param validLen Number of leading token rows to decode.
  * @param grammar Pre-computed grammar tables from {@link buildGrammar}.
@@ -365,6 +367,7 @@ export function viterbiDecode(
 /**
  * Maps a BIOES label id to its bare entity type (the part after the `-`), or
  * an empty string for `O` / unprefixed / out-of-range ids.
+ * @category NLP / Functions
  * @param labelId Predicted label id.
  * @param labelNames The BIOES label list.
  * @returns The entity type, or `''` for background.
@@ -404,6 +407,7 @@ export interface TokenSpan {
  * following same-entity tokens, but stops before the next opener (`B-x`/`S-x`)
  * so two adjacent same-type entities (e.g. `S-x` then `B-x E-x`) stay separate
  * rather than merging into one run.
+ * @category NLP / Functions
  * @param predictedLabels Per-token predicted label ids.
  * @param labelNames The BIOES label list.
  * @returns The detected token spans in order.
@@ -449,6 +453,7 @@ export function extractSpans(
  * Runs `ids.length` `decode` calls, each on a growing prefix. Cheap for the
  * ~256-token windows the privacy filter operates on; keep in mind for larger
  * sequences.
+ * @category NLP / Functions
  * @param tokenizer Same tokenizer instance used to produce `ids`.
  * @param ids Token ids for the whole input.
  * @returns A flat `[start0, end0, start1, end1, ...]` `Uint32Array` of length

From 9733a07189bee82228297c84069b46de4876788c Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:31:30 +0200
Subject: [PATCH 28/38] docs: move LLM above NLP in category order, add
 namespace links to module comment

---
 docs/typedoc.json                             |  4 ++--
 packages/react-native-executorch/src/index.ts | 12 ++++++------
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/docs/typedoc.json b/docs/typedoc.json
index c0339a4683..6a14a31178 100644
--- a/docs/typedoc.json
+++ b/docs/typedoc.json
@@ -16,14 +16,14 @@
     "CV / Tasks",
     "CV / Types",
     "CV / Constants",
+    "LLM / Tasks",
+    "LLM / Types",
     "NLP / Tasks",
     "NLP / Types",
     "NLP / Constants",
     "Speech / Tasks",
     "Speech / Types",
     "Speech / Constants",
-    "LLM / Tasks",
-    "LLM / Types",
     "*"
   ]
 }
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index d6fccf27ed..1f28830b44 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -20,15 +20,15 @@
  *
  * - **Domain Utilities (`cv`, `llm`, `speech`, `nlp`, `math`)**:
  *   Domain-specific low-level tools:
- *   - `cv`: Image transformations (resizing, normalization, layout conversions),
+ *   - {@link math}: Native C++ tensor operations (softmax, sigmoid, argmax, linear,
+ *     activations).
+ *   - {@link cv}: Image transformations (resizing, normalization, layout conversions),
  *     bounding box, quadrilateral, and keypoint helpers.
- *   - `llm`: Token-by-token text generation runner, chat template formatting,
+ *   - {@link llm}: Token-by-token text generation runner, chat template formatting,
  *     and function calling helpers.
- *   - `nlp`: Fast native HuggingFace tokenizers and privacy filter utilities.
- *   - `speech`: Text-to-speech phonemizers, sentence splitters, voice activity
+ *   - {@link nlp}: Fast native HuggingFace tokenizers and privacy filter utilities.
+ *   - {@link speech}: Text-to-speech phonemizers, sentence splitters, voice activity
  *     detection, and audio utilities.
- *   - `math`: Native C++ tensor operations (softmax, sigmoid, argmax, linear,
- *     activations).
  *
  * - **Model Registry ({@link models})**:
  *   A curated catalog of verified, hosted on-device AI models across LLMs,

From cad2b7311691d14aedf2509f72942e8f76bc539a Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:34:38 +0200
Subject: [PATCH 29/38] style

---
 packages/react-native-executorch/src/index.ts | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 1f28830b44..98cef8427b 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -18,7 +18,7 @@
  *   postprocessing, and tokenizers. Ideal for background services, worklets, or
  *   apps needing manual lifecycle control.
  *
- * - **Domain Utilities (`cv`, `llm`, `speech`, `nlp`, `math`)**:
+ * - **Domain Utilities**:
  *   Domain-specific low-level tools:
  *   - {@link math}: Native C++ tensor operations (softmax, sigmoid, argmax, linear,
  *     activations).
@@ -97,6 +97,7 @@ export * from './core/error';
 export * from './core/model';
 export * from './core/tensor';
 export * from './core/runtime';
+
 /**
  * Model schema validation and dimension constraints.
  * @category Modules

From 0a50750f24c8b03a93fda07092a64a33abae832c Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:36:10 +0200
Subject: [PATCH 30/38] docs: add @category tags to imagePreprocessor and
 paddleOcrUtils cv/utils

---
 docs/typedoc.json                                             | 4 ++--
 .../src/extensions/cv/utils/imagePreprocessor.ts              | 3 +++
 .../src/extensions/cv/utils/paddleOcrUtils.ts                 | 2 ++
 3 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/docs/typedoc.json b/docs/typedoc.json
index 6a14a31178..9d7f9dae82 100644
--- a/docs/typedoc.json
+++ b/docs/typedoc.json
@@ -11,8 +11,6 @@
     "Modules",
     "Core / Functions",
     "Core / Types",
-    "Utils / Functions",
-    "Utils / Types",
     "CV / Tasks",
     "CV / Types",
     "CV / Constants",
@@ -24,6 +22,8 @@
     "Speech / Tasks",
     "Speech / Types",
     "Speech / Constants",
+    "Utils / Functions",
+    "Utils / Types",
     "*"
   ]
 }
diff --git a/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts b/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
index 45dc5042d4..e2a5a75a80 100644
--- a/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/imagePreprocessor.ts
@@ -20,6 +20,7 @@ import { RnExecuTorchError } from '../../../core/error';
 
 /**
  * Options for configuring the image preprocessor pipeline.
+ * @category CV / Types
  */
 export type ImagePreprocessorOptions = {
   /** How the input image is resized to match the model's expected dimensions. */
@@ -34,6 +35,7 @@ export type ImagePreprocessorOptions = {
 
 /**
  * Image preprocessor runner for transforming image buffers into model input tensors.
+ * @category CV / Types
  */
 export type ImagePreprocessor = {
   /**
@@ -62,6 +64,7 @@ export type ImagePreprocessor = {
  * normalize, and copy raw image buffers into target tensors matching model
  * input shapes. All intermediate scratch tensors are pre-allocated and safely
  * disposed of when calling `dispose()`.
+ * @category CV / Functions
  * @param options Normalization scaling coefficients, interpolation algorithms, and
  * resize modes.
  * See {@link ImagePreprocessorOptions}.
diff --git a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
index 89468ab202..388cf065dd 100644
--- a/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
+++ b/packages/react-native-executorch/src/extensions/cv/utils/paddleOcrUtils.ts
@@ -30,6 +30,7 @@ function quadsFromFlat(flat: ArrayLike<number>): Quad[] {
 
 /**
  * Thresholds for {@link extractDbnetTextQuads}.
+ * @category CV / Types
  */
 export type DbnetDecodeOptions = {
   /** Binarization threshold on the probability map. */
@@ -48,6 +49,7 @@ export type DbnetDecodeOptions = {
  * Decodes a DBNet probability map into oriented text quads: binarizes the map,
  * traces contours, scores each candidate by its mean probability and unclips the
  * survivors back to their unshrunk size.
+ * @category CV / Functions
  * @param probabilityMap The `detect` output, shape `[1, 1, H, W]`, post-sigmoid.
  * @param options Decode thresholds. See {@link DbnetDecodeOptions}.
  * @returns The decoded quads, in detector-input pixel space and arbitrary order.

From 805e6ca6e2224fb3d39f82a504d0fc89bd4aaaf8 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:42:32 +0200
Subject: [PATCH 31/38] style

---
 packages/react-native-executorch/src/index.ts | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 98cef8427b..12a9b13271 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -5,18 +5,14 @@
  * It provides a layered architecture designed for different use cases:
  *
  * - **React Hooks (`use<Task>`)**:
- *   Declarative hooks (e.g. {@link useClassifier}, {@link useLLMChatSession},
- *   {@link useSpeechToText}, {@link useOpticalCharacterRecognizer}) designed
- *   for UI components. They manage downloading remote model assets with
- *   progress tracking, compilation, and automatic native memory disposal on
- *   unmount.
+ *   Declarative hooks designed for UI components. They manage downloading
+ *   remote model assets with progress tracking, compilation, and automatic
+ *   native memory disposal on unmount.
  *
  * - **Task APIs (`create<Task>`)**:
- *   Imperative, promise-based pipelines (e.g. {@link createClassifier},
- *   {@link createLLMChatSession}, {@link createPaddleOcr},
- *   {@link createWhisperSpeechToText}) with built-in preprocessing,
- *   postprocessing, and tokenizers. Ideal for background services, worklets, or
- *   apps needing manual lifecycle control.
+ *   Imperative, promise-based pipelines with built-in preprocessing and
+ *   postprocessing. Ideal for background services, worklets, or apps needing
+ *   manual lifecycle control.
  *
  * - **Domain Utilities**:
  *   Domain-specific low-level tools:

From 794065e580898bc1f8fdb83b84044dcbbac7ed49 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 15:52:34 +0200
Subject: [PATCH 32/38] style

---
 packages/react-native-executorch/src/index.ts | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 12a9b13271..f6bad4674a 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -16,8 +16,8 @@
  *
  * - **Domain Utilities**:
  *   Domain-specific low-level tools:
- *   - {@link math}: Native C++ tensor operations (softmax, sigmoid, argmax, linear,
- *     activations).
+ *   - {@link math}: Native C++ tensor operations (softmax, sigmoid, argmax,
+ *     gather, etc.).
  *   - {@link cv}: Image transformations (resizing, normalization, layout conversions),
  *     bounding box, quadrilateral, and keypoint helpers.
  *   - {@link llm}: Token-by-token text generation runner, chat template formatting,

From 3826726a07d2f42edaa1d11a9843fe64e1bcb1dd Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 16:30:20 +0200
Subject: [PATCH 33/38] refactor(schema): rename constr -> constraint, eq ->
 equality

---
 .../src/core/schema.ts                        | 24 +++++++++----------
 .../src/extensions/cv/tasks/paddleOcr.ts      |  4 ++--
 .../src/extensions/nlp/tasks/privacyFilter.ts | 11 +++++++--
 .../src/extensions/nlp/tasks/textEmbedding.ts | 13 +++++++---
 .../speech/tasks/kokoroTextToSpeech.ts        |  8 +++----
 .../speech/tasks/supertonicTextToSpeech.ts    | 19 ++++++++++-----
 6 files changed, 50 insertions(+), 29 deletions(-)

diff --git a/packages/react-native-executorch/src/core/schema.ts b/packages/react-native-executorch/src/core/schema.ts
index 7ae0ba4aa4..6e74bbb62e 100644
--- a/packages/react-native-executorch/src/core/schema.ts
+++ b/packages/react-native-executorch/src/core/schema.ts
@@ -376,13 +376,13 @@ export const bool = (...shape: SymbolicShape) => SymbolicTensor('bool', shape);
  * Helper namespace for declaring runtime constraints.
  * @category Core / Schema / Functions
  */
-export const constr = {
+export const constraint = {
   /**
    * Declares that all given dimensions must be equal at runtime.
    * @param dims Dimensions that must share the same concrete value.
    * @returns An {@link EqualityConstraint} across the given dimensions.
    */
-  eq: (...dims: DimRef[]): EqualityConstraint => {
+  equality: (...dims: DimRef[]): EqualityConstraint => {
     return { kind: 'equality', dims };
   },
   /**
@@ -647,8 +647,8 @@ function matchRuntimeConstraints(
 
     const unclaimed = [...exportedMethodSpec.runtimeConstraints];
 
-    for (const [idx, constraint] of allowedMethodSpec.runtimeConstraints.entries()) {
-      const find = unclaimed.findIndex((c) => constraintsEqual(c, constraint));
+    for (const [idx, rc] of allowedMethodSpec.runtimeConstraints.entries()) {
+      const find = unclaimed.findIndex((c) => constraintsEqual(c, rc));
       if (find === -1) {
         throw RnExecuTorchError(
           'SCHEMA_MISMATCH',
@@ -696,26 +696,26 @@ function validateSymbolKindConsistency(modelSpec: ModelSpec<SymbolicDim>): void
 
 function validateConstraintCorrectness(modelSpec: ModelSpec<SymbolicDim>): void {
   for (const [methodName, methodSpec] of Object.entries(modelSpec)) {
-    for (const [idx, constraint] of methodSpec.runtimeConstraints.entries()) {
+    for (const [idx, rc] of methodSpec.runtimeConstraints.entries()) {
       const ctx = `Method '${methodName}' constraint ${idx}`;
 
-      if (constraint.kind === 'linear') {
-        const [A, B] = constraint.coefficients;
+      if (rc.kind === 'linear') {
+        const [A, B] = rc.coefficients;
         if (!Number.isInteger(A) || !Number.isInteger(B)) {
           throw RnExecuTorchError('SCHEMA_MISMATCH', `${ctx}: Coefficients must be integers.`);
         }
-        resolveDim(methodSpec, constraint.dimLhs);
-        resolveDim(methodSpec, constraint.dimRhs);
+        resolveDim(methodSpec, rc.dimLhs);
+        resolveDim(methodSpec, rc.dimRhs);
       }
 
-      if (constraint.kind === 'equality') {
-        if (constraint.dims.length < 2) {
+      if (rc.kind === 'equality') {
+        if (rc.dims.length < 2) {
           throw RnExecuTorchError(
             'SCHEMA_MISMATCH',
             `${ctx}: Equality requires at least two dimensions.`
           );
         }
-        constraint.dims.forEach((ref) => resolveDim(methodSpec, ref));
+        rc.dims.forEach((ref) => resolveDim(methodSpec, ref));
       }
     }
   }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 5472e559e5..89b595ed85 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -13,7 +13,7 @@ import { tensor, type Tensor } from '../../../core/tensor';
 import {
   validateSpec,
   method,
-  constr,
+  constraint,
   f32,
   DynamicDim,
   type ConcreteDim,
@@ -418,7 +418,7 @@ export async function createPaddleOcr(
         [f32(1, 3, 'recH', DynamicDim('recW'))],
         [f32(1, DynamicDim('recT'), 'vocab')],
         [
-          constr.linear(
+          constraint.linear(
             { paramSide: 'input', tensorIdx: 0, dimIdx: 3 },
             { paramSide: 'output', tensorIdx: 0, dimIdx: 1 },
             SVTR_CTC_STRIDE
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
index ed2482b2a3..7230f5581f 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/privacyFilter.ts
@@ -8,7 +8,14 @@ import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
 import { loadModel } from '../../../core/model';
-import { validateSpec, DynamicDim as Dyn, method, i64, f32, constr } from '../../../core/schema';
+import {
+  validateSpec,
+  DynamicDim as Dyn,
+  method,
+  i64,
+  f32,
+  constraint,
+} from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 import { RnExecuTorchError } from '../../../core/error';
 
@@ -146,7 +153,7 @@ export async function createPrivacyFilter<Label extends string>(
   // it dynamic, binding `S` to the accepted range; the MLX ones are exported
   // for a single window length, binding `S` to that constant.
   const seqLenIsShared = [
-    constr.eq(
+    constraint.equality(
       { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
       { paramSide: 'input', tensorIdx: 1, dimIdx: 1 },
       { paramSide: 'output', tensorIdx: 0, dimIdx: 1 }
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
index d5c828b1af..07111f0943 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
@@ -6,7 +6,14 @@ import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor } from '../../../core/tensor';
 import { loadModel } from '../../../core/model';
-import { validateSpec, DynamicDim as Dyn, method, i64, f32, constr } from '../../../core/schema';
+import {
+  validateSpec,
+  DynamicDim as Dyn,
+  method,
+  i64,
+  f32,
+  constraint,
+} from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 
 import { loadTokenizer } from '../tokenizer';
@@ -94,7 +101,7 @@ export async function createTextEmbedder(
       [i64(1, Dyn('L')), i64(1, Dyn('L'))],
       [f32(1, 'D')],
       [
-        constr.eq(
+        constraint.equality(
           { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
           { paramSide: 'input', tensorIdx: 1, dimIdx: 1 }
         ),
@@ -105,7 +112,7 @@ export async function createTextEmbedder(
       [i64(1, Dyn('L')), i64(1, Dyn('L'))],
       [f32('D')],
       [
-        constr.eq(
+        constraint.equality(
           { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
           { paramSide: 'input', tensorIdx: 1, dimIdx: 1 }
         ),
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
index 70059a214c..098a991c13 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/kokoroTextToSpeech.ts
@@ -15,7 +15,7 @@ import {
   f32,
   bool,
   DynamicDim as Dyn,
-  constr,
+  constraint,
 } from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 import { RnExecuTorchError } from '../../../core/error';
@@ -179,7 +179,7 @@ export async function createKokoroTextToSpeech<K extends PropertyKey>(
           f32(1, Dyn('T'), DURATION_FEATURE_DIM), // durationFeatures
         ],
         [
-          constr.eq(
+          constraint.equality(
             { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
             { paramSide: 'input', tensorIdx: 1, dimIdx: 1 },
             { paramSide: 'output', tensorIdx: 0, dimIdx: 0 },
@@ -201,12 +201,12 @@ export async function createKokoroTextToSpeech<K extends PropertyKey>(
         ],
         [f32(1, 1, Dyn('AUDIO_LEN'))], // audio
         [
-          constr.eq(
+          constraint.equality(
             { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
             { paramSide: 'input', tensorIdx: 1, dimIdx: 1 },
             { paramSide: 'input', tensorIdx: 3, dimIdx: 1 }
           ),
-          constr.linear(
+          constraint.linear(
             { paramSide: 'output', tensorIdx: 0, dimIdx: 2 },
             { paramSide: 'input', tensorIdx: 2, dimIdx: 0 },
             TICKS_PER_DURATION
diff --git a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
index 903b5013f7..647b058df9 100644
--- a/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
+++ b/packages/react-native-executorch/src/extensions/speech/tasks/supertonicTextToSpeech.ts
@@ -9,7 +9,14 @@ import RNBlobUtil from 'react-native-blob-util';
 
 import { tensor, type Tensor } from '../../../core/tensor';
 import { loadModel } from '../../../core/model';
-import { validateSpec, method, i64, f32, DynamicDim as Dyn, constr } from '../../../core/schema';
+import {
+  validateSpec,
+  method,
+  i64,
+  f32,
+  DynamicDim as Dyn,
+  constraint,
+} from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 import { randomNormal } from '../../math';
 import {
@@ -192,7 +199,7 @@ export async function createSupertonicTextToSpeech<K extends PropertyKey>(
       ],
       [f32(1)], // duration
       [
-        constr.eq(
+        constraint.equality(
           { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
           { paramSide: 'input', tensorIdx: 2, dimIdx: 2 }
         ),
@@ -210,7 +217,7 @@ export async function createSupertonicTextToSpeech<K extends PropertyKey>(
       ],
       [f32(1, TEXT_EMB_DIM, Dyn('T'))], // textEmb
       [
-        constr.eq(
+        constraint.equality(
           { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
           { paramSide: 'input', tensorIdx: 2, dimIdx: 2 },
           { paramSide: 'output', tensorIdx: 0, dimIdx: 2 }
@@ -233,12 +240,12 @@ export async function createSupertonicTextToSpeech<K extends PropertyKey>(
       ],
       [f32(1, LATENT_DIM, Dyn('L'))], // denoisedLatent
       [
-        constr.eq(
+        constraint.equality(
           { paramSide: 'input', tensorIdx: 0, dimIdx: 2 },
           { paramSide: 'input', tensorIdx: 4, dimIdx: 2 },
           { paramSide: 'output', tensorIdx: 0, dimIdx: 2 }
         ),
-        constr.eq(
+        constraint.equality(
           { paramSide: 'input', tensorIdx: 1, dimIdx: 2 },
           { paramSide: 'input', tensorIdx: 3, dimIdx: 2 }
         ),
@@ -252,7 +259,7 @@ export async function createSupertonicTextToSpeech<K extends PropertyKey>(
       [f32(1, LATENT_DIM, Dyn('L'))], // latent
       [f32(1, Dyn('AUDIO_LEN'))], // wav
       [
-        constr.linear(
+        constraint.linear(
           { paramSide: 'output', tensorIdx: 0, dimIdx: 1 },
           { paramSide: 'input', tensorIdx: 0, dimIdx: 2 },
           CHUNK_SIZE

From 736ffd6ca171fa82ebc976796815fe53e561628f Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 18:17:05 +0200
Subject: [PATCH 34/38] fix: update JSDoc references to use proper {/@link}
 syntax

- Replace modelOpts references with linked option types in instanceSegmentation, objectDetection, and keypointDetection
- Add links to boundingBoxOfPoints, TextEmbedder.embed, Tokenizer.decode, and download functions
- Preserve original comments while adding proper documentation links
---
 .../src/extensions/cv/tasks/instanceSegmentation.ts         | 6 +++---
 .../src/extensions/cv/tasks/keypointDetection.ts            | 4 ++--
 .../src/extensions/cv/tasks/objectDetection.ts              | 4 ++--
 .../src/extensions/cv/tasks/paddleOcr.ts                    | 2 +-
 .../src/extensions/nlp/utils/privacyFilterUtils.ts          | 3 ++-
 packages/react-native-executorch/src/models.ts              | 2 +-
 packages/react-native-executorch/src/utils.ts               | 2 +-
 7 files changed, 12 insertions(+), 11 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
index d4bc55fe0a..1390ab185c 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/instanceSegmentation.ts
@@ -73,17 +73,17 @@ export type InstanceSegmenterModel<F extends BoxFormat, L> = {
 export type SegmentInstancesOptions = {
   /**
    * Minimum confidence threshold. If omitted, uses
-   * `modelOpts.defaultConfidenceThreshold`.
+   * {@link InstanceSegmenterOptions.defaultConfidenceThreshold}.
    */
   readonly confidenceThreshold?: number;
   /**
    * Intersection over Union (IoU) threshold in NMS. If omitted, uses
-   * `modelOpts.defaultIouThreshold`.
+   * {@link InstanceSegmenterOptions.defaultIouThreshold}.
    */
   readonly iouThreshold?: number;
   /**
    * Mask binarization probability threshold. If omitted, uses
-   * `modelOpts.defaultMaskThreshold`.
+   * {@link InstanceSegmenterOptions.defaultMaskThreshold}.
    */
   readonly maskThreshold?: number;
 };
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
index 2cd7771610..7b168625a4 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/keypointDetection.ts
@@ -59,12 +59,12 @@ export type KeypointDetectorModel<F extends BoxFormat, L extends PropertyKey> =
 export type DetectKeypointsOptions = {
   /**
    * Minimum confidence score threshold for detections. If omitted, uses
-   * `modelOpts.defaultConfidenceThreshold`.
+   * {@link KeypointDetectorOptions.defaultConfidenceThreshold}.
    */
   readonly confidenceThreshold?: number;
   /**
    * Intersection over Union (IoU) threshold for NMS. If omitted, uses
-   * `modelOpts.defaultIouThreshold`.
+   * {@link KeypointDetectorOptions.defaultIouThreshold}.
    */
   readonly iouThreshold?: number;
 };
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
index 7760ce05fa..833890527a 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/objectDetection.ts
@@ -59,12 +59,12 @@ export type ObjectDetectorModel<F extends BoxFormat, L> = {
 export type DetectObjectsOptions = {
   /**
    * Minimum confidence score threshold. If omitted, uses
-   * `modelOpts.defaultConfidenceThreshold`.
+   * {@link ObjectDetectorOptions.defaultConfidenceThreshold}.
    */
   readonly confidenceThreshold?: number;
   /**
    * Intersection over Union (IoU) threshold for NMS. If omitted, uses
-   * `modelOpts.defaultIouThreshold`.
+   * {@link ObjectDetectorOptions.defaultIouThreshold}.
    */
   readonly iouThreshold?: number;
 };
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
index 89b595ed85..6d5f4bf7cf 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/paddleOcr.ts
@@ -53,7 +53,7 @@ export type OcrDetection = {
   /**
    * The region's corners, ordered top-left, top-right, bottom-right, bottom-left,
    * in original image pixels. Oriented, so a rotated line keeps its angle; take
-   * `boundingBoxOfPoints(quad, 'xyxy')` for the axis-aligned box.
+   * {@link boundingBoxOfPoints} for the axis-aligned box.
    */
   readonly quad: Quad;
 };
diff --git a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
index c0384f081d..3c3f317290 100644
--- a/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/utils/privacyFilterUtils.ts
@@ -448,7 +448,8 @@ export function extractSpans(
  * Assumes a byte-level BPE tokenizer with no lossy normalization (the openai
  * and nemotron privacy filter presets satisfy this): decoding the full token
  * sequence must reproduce the input string exactly. Under that guarantee, the
- * length of `decode(ids[0..k])` is the character offset just past token `k-1`.
+ * length of `decode(ids[0..k])` (see: {@link Tokenizer.decode}) is the
+ * character offset just past token `k-1`.
  *
  * Runs `ids.length` `decode` calls, each on a growing prefix. Cheap for the
  * ~256-token windows the privacy filter operates on; keep in mind for larger
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 2f0bfafea8..4b653a4732 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -2167,7 +2167,7 @@ export const models = {
     /**
      * Liquid AI LFM 2.5 350M parameter embedding model for asymmetric search
      * and retrieval tasks. Prompts queries with `query: ` (the default) and
-     * passages with `document: ` via `embed(text, 'document: ')`.
+     * passages with `document: ` via {@link TextEmbedder.embed}.
      */
     LFM2_5_EMBEDDING_350M: {
       DEFAULT: LFM2_5_EMBEDDING_350M_EMBEDDINGS,
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index aed28dfe95..49f8944fdf 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -74,7 +74,7 @@ export async function inspectModel(source: string): Promise<ModelInspection> {
 
   if (source.startsWith('http')) {
     // Throwaway download to a temp path — inspection shouldn't populate the
-    // persistent resource cache, so we don't go through `download()`.
+    // persistent resource cache, so we don't go through {@link download}.
     localPath = `${RNBlobUtil.fs.dirs.CacheDir}/inspect_model_${Date.now()}.pte`;
     await RNBlobUtil.config({ path: localPath }).fetch('GET', source);
     downloaded = true;

From c19d3395a8f5eaf5ff673ed429561eae3e9d0f4b Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 18:18:16 +0200
Subject: [PATCH 35/38] Add missing readonly

---
 packages/react-native-executorch/src/utils.ts | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index 49f8944fdf..2c495c914c 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -15,7 +15,7 @@ import RNBlobUtil from 'react-native-blob-util';
  * @returns An array of registered backend name strings (e.g. 'XnnpackBackend',
  * 'CoreMLBackend').
  */
-export function getRegisteredBackends(): string[] {
+export function getRegisteredBackends(): readonly string[] {
   'worklet';
   return rnexecutorchJsi.getExecuTorchRegisteredBackends();
 }

From 72325533f1a78ded2f91517f87212473c8f735f5 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Tue, 25 Aug 2026 18:28:16 +0200
Subject: [PATCH 36/38] fix: add Core / Constants category and fix missing
 readonly and links

- Add Core / Constants to VALID_CATEGORIES in .eslintrc.js
- Add Core / Constants to categoryOrder in typedoc.json
- Move defaultWorkletRuntime to Core / Constants category
- Move VALID_ERROR_CODES to Core / Constants category
- Add missing readonly modifiers to ResourceOptions type
- Add proper link to TextEmbedderModel.defaultPrompt in textEmbedding.ts
---
 .eslintrc.js                                          |  1 +
 docs/typedoc.json                                     | 11 ++++++-----
 packages/react-native-executorch/src/core/error.ts    |  2 +-
 packages/react-native-executorch/src/core/runtime.ts  |  2 +-
 .../src/extensions/nlp/tasks/textEmbedding.ts         |  2 +-
 packages/react-native-executorch/src/utils.ts         |  4 ++--
 6 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/.eslintrc.js b/.eslintrc.js
index f2e119957a..9b4d7f485e 100644
--- a/.eslintrc.js
+++ b/.eslintrc.js
@@ -5,6 +5,7 @@ const VALID_CATEGORIES = [
   'CV / Functions',
   'CV / Tasks',
   'CV / Types',
+  'Core / Constants',
   'Core / Errors',
   'Core / Functions',
   'Core / Schema / Functions',
diff --git a/docs/typedoc.json b/docs/typedoc.json
index 9d7f9dae82..4361f2f06c 100644
--- a/docs/typedoc.json
+++ b/docs/typedoc.json
@@ -11,16 +11,17 @@
     "Modules",
     "Core / Functions",
     "Core / Types",
+    "Core / Constants",
     "CV / Tasks",
-    "CV / Types",
-    "CV / Constants",
     "LLM / Tasks",
-    "LLM / Types",
     "NLP / Tasks",
-    "NLP / Types",
-    "NLP / Constants",
     "Speech / Tasks",
+    "CV / Types",
+    "LLM / Types",
+    "NLP / Types",
     "Speech / Types",
+    "CV / Constants",
+    "NLP / Constants",
     "Speech / Constants",
     "Utils / Functions",
     "Utils / Types",
diff --git a/packages/react-native-executorch/src/core/error.ts b/packages/react-native-executorch/src/core/error.ts
index 84374a6967..952f290b01 100644
--- a/packages/react-native-executorch/src/core/error.ts
+++ b/packages/react-native-executorch/src/core/error.ts
@@ -18,7 +18,7 @@
  * resource, re-create a disposed one). Everything else is a category that
  * exists so crash reporters can group failures, and the detail lives in the
  * message.
- * @category Core / Types
+ * @category Core / Constants
  */
 export const VALID_ERROR_CODES = [
   'LOAD_FAILED',
diff --git a/packages/react-native-executorch/src/core/runtime.ts b/packages/react-native-executorch/src/core/runtime.ts
index 86b1bb62e8..6c68dcc340 100644
--- a/packages/react-native-executorch/src/core/runtime.ts
+++ b/packages/react-native-executorch/src/core/runtime.ts
@@ -20,7 +20,7 @@ import { isRnExecuTorchError, RnExecuTorchError } from './error';
  * thread, preventing model loading and inference from blocking the UI. Pass it
  * explicitly (or a custom {@link WorkletRuntime}) to {@link wrapAsync} when you
  * need fine-grained control over which thread work executes on.
- * @category Core / Functions
+ * @category Core / Constants
  */
 export const defaultWorkletRuntime = createWorkletRuntime({
   name: 'ExecuTorchDefaultRuntime',
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
index 07111f0943..b619dd0c8f 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/textEmbedding.ts
@@ -47,7 +47,7 @@ export type TextEmbedder = {
    * Inputs longer than the model's maximum sequence length are truncated.
    * @param input The input text to embed.
    * @param prompt Optional prompt prefix overriding the model's configured
-   * `defaultPrompt` for this call.
+   * {@link TextEmbedderModel.defaultPrompt} for this call.
    * @returns A promise resolving to the embedding vector.
    * @throws {RnExecuTorchError} With code `INVALID_ARGUMENT` if the input text
    * tokenizes to zero tokens, `RESOURCE_BUSY` if the model is in use, or
diff --git a/packages/react-native-executorch/src/utils.ts b/packages/react-native-executorch/src/utils.ts
index 2c495c914c..1bb5830404 100644
--- a/packages/react-native-executorch/src/utils.ts
+++ b/packages/react-native-executorch/src/utils.ts
@@ -27,13 +27,13 @@ export function getRegisteredBackends(): readonly string[] {
  */
 export type ResourceOptions = {
   /** If true, prevents checks and downloads, resetting the hook state. */
-  preventLoad?: boolean;
+  readonly preventLoad?: boolean;
   /**
    * Re-downloads every remote source even when it is already cached, replacing
    * the cached copy. Use to recover from a corrupted file or to pick up a model
    * that changed behind a stable URL.
    */
-  forceDownload?: boolean;
+  readonly forceDownload?: boolean;
 };
 
 /**

From 79483e00d9588260ba7269bdec60e85f76b779ad Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Wed, 26 Aug 2026 15:08:13 +0200
Subject: [PATCH 37/38] fix: update agent skills, typedoc categories, model
 defaults, and docs warnings

---
 .agents/skills/add-task-pipeline/SKILL.md     |   2 +-
 .../skills/model-schema-validation/SKILL.md   |  20 ++--
 .eslintrc.js                                  |  32 +-----
 docs/typedoc.json                             |  18 ++-
 .../src/extensions/nlp/tasks/tokenization.ts  |   2 +-
 .../src/hooks/useTextToSpeech.ts              |  51 +++++----
 .../react-native-executorch/src/models.ts     | 105 +++++++++++++++---
 7 files changed, 142 insertions(+), 88 deletions(-)

diff --git a/.agents/skills/add-task-pipeline/SKILL.md b/.agents/skills/add-task-pipeline/SKILL.md
index d0d9d306a6..6eefdabd72 100644
--- a/.agents/skills/add-task-pipeline/SKILL.md
+++ b/.agents/skills/add-task-pipeline/SKILL.md
@@ -100,7 +100,7 @@ import { loadModel } from '../../../core/model';
 import { validateSpec, method, f32 } from '../../../core/schema';
 import { wrapAsync } from '../../../core/runtime';
 import { type ImageBuffer } from '../image';
-import { createImagePreprocessor, type ImagePreprocessorOptions } from './preprocessing';
+import { createImagePreprocessor, type ImagePreprocessorOptions } from '../utils/imagePreprocessor';
 
 export type MyTaskOptions = ImagePreprocessorOptions & {
   readonly defaultThreshold: number;
diff --git a/.agents/skills/model-schema-validation/SKILL.md b/.agents/skills/model-schema-validation/SKILL.md
index 9b6cfd2254..9b39dd6106 100644
--- a/.agents/skills/model-schema-validation/SKILL.md
+++ b/.agents/skills/model-schema-validation/SKILL.md
@@ -30,7 +30,7 @@ import {
   i64,
   i32,
   DynamicDim as Dyn,
-  constr,
+  constraint,
 } from '../../../core/schema';
 
 const { variant, dims } = validateSpec(model.schema, {
@@ -39,7 +39,7 @@ const { variant, dims } = validateSpec(model.schema, {
     [i64(1, Dyn('L')), i64(1, Dyn('L'))],
     [f32(1, 'D')],
     [
-      constr.eq(
+      constraint.equality(
         { paramSide: 'input', tensorIdx: 0, dimIdx: 1 },
         { paramSide: 'input', tensorIdx: 1, dimIdx: 1 }
       ),
@@ -50,7 +50,7 @@ const { variant, dims } = validateSpec(model.schema, {
     [i64(Dyn('L')), i64(Dyn('L'))],
     [f32('D')],
     [
-      constr.eq(
+      constraint.equality(
         { paramSide: 'input', tensorIdx: 0, dimIdx: 0 },
         { paramSide: 'input', tensorIdx: 1, dimIdx: 0 }
       ),
@@ -68,10 +68,10 @@ const L = dims.range('L');
 - **`f32(...)` / `i64(...)` / `i32(...)` / `ui8(...)`**: Shorthand helpers for tensor parameter specs.
 - **`StaticDim('symbol')` / String Literals**: Strings passed to shape helpers (e.g. `'H'`, `'W'`) automatically map to `StaticDim`, acting as **static dimension wildcards**. They bind strictly to `constant` positive integer dimensions in the exported spec.
 - **`DynamicDim('symbol')` (or `Dyn('symbol')`)**: Creates a dynamic dimension symbol. Must be used when a dimension genuinely varies at runtime and binds to a `range` or `enum` domain in the exported spec.
-- **Constraint Helpers (`constr`)**:
+- **Constraint Helpers (`constraint`)**:
   - **`DimRef` Object Literal (`{ paramSide: 'input' | 'output', tensorIdx, dimIdx }`)**: Explicit reference to a tensor's dimension.
-  - **`constr.eq(...dims)`**: Creates an equality constraint requiring the referenced dimensions to take the exact same value at runtime.
-  - **`constr.linear(lhs, rhs, a, b)`**: Creates a linear constraint `lhs = a * rhs + b`.
+  - **`constraint.equality(...dims)`**: Creates an equality constraint requiring the referenced dimensions to take the exact same value at runtime.
+  - **`constraint.linear(lhs, rhs, a, b?)`**: Creates a linear constraint `lhs = a * rhs + b`.
 - **`validateSpec(exportedSchema, allowedVariants)`**: Compares the model's exported schema against named variants and returns `{ variant, dim, dims }`.
 - **Symbol Accessors (`dims` & `dim`)**:
   - `dims.constant('N', 'H')`: Extracts constant values for symbols as numbers.
@@ -118,24 +118,24 @@ Understanding the distinction between a dimension's **domain** and its **runtime
   - Dynamic symbols (`DynamicDim('S')`) bind to exported dimension domains. Reusing a symbol (`'S'`) across tensor inputs or outputs requires every occurrence to bind to the **exact same domain**.
   - ⚠️ **Key Rule**: Binding to the same domain does **NOT** mean runtime values coincide! Two dimensions bound to the same domain (e.g., both having range `1..512`) may take _different_ runtime values in a single execution (e.g. length 10 and length 25).
 
-### 2. Runtime Constraints (`constr.eq` & `constr.linear`)
+### 2. Runtime Constraints (`constraint.equality` & `constraint.linear`)
 
 - **Runtime Constraints**: Declarations about the **runtime values** of tensor dimensions during execution:
-  - **Equality Constraint (`constr.eq(...)`)**: Requires all referenced dimensions to take the exact same runtime value in any execution call.
+  - **Equality Constraint (`constraint.equality(...)`)**: Requires all referenced dimensions to take the exact same runtime value in any execution call.
     ```typescript
     method(
       'forward',
       [f32('B', Dyn('S1')), f32('B', Dyn('S2'))],
       [f32('B', Dyn('S1'))],
       [
-        constr.eq(
+        constraint.equality(
           { paramSide: 'input', tensorIdx: 0, dimIdx: 0 },
           { paramSide: 'input', tensorIdx: 1, dimIdx: 0 }
         ),
       ]
     );
     ```
-  - **Linear Constraint (`constr.linear(...)`)**: Requires two dimensions to satisfy `dimLhs = a * dimRhs + b` at runtime.
+  - **Linear Constraint (`constraint.linear(...)`)**: Requires two dimensions to satisfy `dimLhs = a * dimRhs + b` at runtime.
 - **Validation & Enforcement**:
   - `validateSpec` verifies that the exported model spec declares the exact same runtime constraints (1-to-1 declaration match).
   - Native C++ validates input runtime constraints before invoking `model.execute()`.
diff --git a/.eslintrc.js b/.eslintrc.js
index 9b4d7f485e..94b518b535 100644
--- a/.eslintrc.js
+++ b/.eslintrc.js
@@ -1,35 +1,7 @@
 const path = require('path');
+const typedocConfig = require('./docs/typedoc.json');
 
-const VALID_CATEGORIES = [
-  'CV / Constants',
-  'CV / Functions',
-  'CV / Tasks',
-  'CV / Types',
-  'Core / Constants',
-  'Core / Errors',
-  'Core / Functions',
-  'Core / Schema / Functions',
-  'Core / Schema / Types',
-  'Core / Types',
-  'Hooks',
-  'LLM / Functions',
-  'LLM / Tasks',
-  'LLM / Types',
-  'Math / Functions',
-  'Math / Types',
-  'Models',
-  'Modules',
-  'NLP / Constants',
-  'NLP / Functions',
-  'NLP / Tasks',
-  'NLP / Types',
-  'Speech / Constants',
-  'Speech / Functions',
-  'Speech / Tasks',
-  'Speech / Types',
-  'Utils / Functions',
-  'Utils / Types',
-];
+const VALID_CATEGORIES = typedocConfig.categoryOrder.filter((cat) => cat !== '*');
 
 const CATEGORY_TAG_MATCH = `^(${VALID_CATEGORIES.join('|')})$`;
 
diff --git a/docs/typedoc.json b/docs/typedoc.json
index 4361f2f06c..503ddc63d0 100644
--- a/docs/typedoc.json
+++ b/docs/typedoc.json
@@ -12,17 +12,25 @@
     "Core / Functions",
     "Core / Types",
     "Core / Constants",
+    "Core / Schema / Functions",
+    "Core / Schema / Types",
     "CV / Tasks",
-    "LLM / Tasks",
-    "NLP / Tasks",
-    "Speech / Tasks",
+    "CV / Functions",
     "CV / Types",
+    "CV / Constants",
+    "LLM / Tasks",
+    "LLM / Functions",
     "LLM / Types",
+    "NLP / Tasks",
+    "NLP / Functions",
     "NLP / Types",
-    "Speech / Types",
-    "CV / Constants",
     "NLP / Constants",
+    "Speech / Tasks",
+    "Speech / Functions",
+    "Speech / Types",
     "Speech / Constants",
+    "Math / Functions",
+    "Math / Types",
     "Utils / Functions",
     "Utils / Types",
     "*"
diff --git a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
index 722e3e0195..22076e179e 100644
--- a/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
+++ b/packages/react-native-executorch/src/extensions/nlp/tasks/tokenization.ts
@@ -13,7 +13,7 @@ import { loadTokenizer } from '../tokenizer';
  * @category NLP / Tasks
  * @param tokenizerPath Absolute local path to a `tokenizer.json` file.
  * @param runtime Optional worklet runtime thread to run the tokenizer on.
- * @returns A promise resolving to the instantiated {@link Tokenizer} operations
+ * @returns A promise resolving to the instantiated tokenizer operations
  * and disposal controls.
  */
 export async function createTokenizer(tokenizerPath: string, runtime?: WorkletRuntime) {
diff --git a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
index 44b0f849e5..4ab8b90fc2 100644
--- a/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
+++ b/packages/react-native-executorch/src/hooks/useTextToSpeech.ts
@@ -11,27 +11,6 @@ import {
   type SupertonicTextToSpeech,
 } from '../extensions/speech/tasks/supertonicTextToSpeech';
 
-type TtsHookResult<
-  C,
-  P extends {
-    synthesize: (...args: any[]) => any;
-    synthesizeStop: (...args: any[]) => any;
-  },
-> = {
-  /** Whether the pipeline is loaded and ready to synthesize. */
-  isReady: boolean;
-  /** The download or load error, if any. */
-  error: Error | undefined;
-  /** Download progress across every asset, in percent. */
-  downloadProgress: number;
-  /** The config with every remote URL resolved to a local path. */
-  resource: C | undefined;
-  /** Streams synthesized audio chunks. Undefined until the pipeline is ready. */
-  synthesize: P['synthesize'] | undefined;
-  /** Cancels an in-flight synthesis. Undefined until the pipeline is ready. */
-  synthesizeStop: P['synthesizeStop'] | undefined;
-};
-
 /**
  * React hook to load and run the Kokoro Text-to-Speech pipeline.
  *
@@ -52,7 +31,20 @@ type TtsHookResult<
 export function useTextToSpeech<K extends PropertyKey>(
   config: KokoroTtsModel<K>,
   options?: ResourceOptions
-): TtsHookResult<KokoroTtsModel<K>, KokoroTextToSpeech<K>>;
+): {
+  /** Whether the pipeline is loaded and ready to synthesize. */
+  isReady: boolean;
+  /** The download or load error, if any. */
+  error: Error | undefined;
+  /** Download progress across every asset, in percent. */
+  downloadProgress: number;
+  /** The config with every remote URL resolved to a local path. */
+  resource: KokoroTtsModel<K> | undefined;
+  /** Streams synthesized audio chunks. Undefined until the pipeline is ready. */
+  synthesize: KokoroTextToSpeech<K>['synthesize'] | undefined;
+  /** Cancels an in-flight synthesis. Undefined until the pipeline is ready. */
+  synthesizeStop: KokoroTextToSpeech<K>['synthesizeStop'] | undefined;
+};
 
 /**
  * React hook to load and run the Supertonic 3 Text-to-Speech pipeline.
@@ -75,7 +67,20 @@ export function useTextToSpeech<K extends PropertyKey>(
 export function useTextToSpeech<K extends PropertyKey>(
   config: SupertonicTtsModel<K>,
   options?: ResourceOptions
-): TtsHookResult<SupertonicTtsModel<K>, SupertonicTextToSpeech<K>>;
+): {
+  /** Whether the pipeline is loaded and ready to synthesize. */
+  isReady: boolean;
+  /** The download or load error, if any. */
+  error: Error | undefined;
+  /** Download progress across every asset, in percent. */
+  downloadProgress: number;
+  /** The config with every remote URL resolved to a local path. */
+  resource: SupertonicTtsModel<K> | undefined;
+  /** Streams synthesized audio chunks. Undefined until the pipeline is ready. */
+  synthesize: SupertonicTextToSpeech<K>['synthesize'] | undefined;
+  /** Cancels an in-flight synthesis. Undefined until the pipeline is ready. */
+  synthesizeStop: SupertonicTextToSpeech<K>['synthesizeStop'] | undefined;
+};
 
 export function useTextToSpeech<K extends PropertyKey>(
   config: KokoroTtsModel<K> | SupertonicTtsModel<K>,
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 4b653a4732..1985444133 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -1531,14 +1531,17 @@ export const models = {
       NANO: {
         DEFAULT: YOLO26_NANO_384_XNNPACK_FP32,
         SIZE_384: {
+          DEFAULT: YOLO26_NANO_384_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_NANO_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_NANO_384_COREML_FP16,
         },
         SIZE_512: {
+          DEFAULT: YOLO26_NANO_512_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_NANO_512_XNNPACK_FP32,
           COREML_FP16: YOLO26_NANO_512_COREML_FP16,
         },
         SIZE_640: {
+          DEFAULT: YOLO26_NANO_640_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_NANO_640_XNNPACK_FP32,
           COREML_FP16: YOLO26_NANO_640_COREML_FP16,
         },
@@ -1550,14 +1553,17 @@ export const models = {
       SMALL: {
         DEFAULT: YOLO26_SMALL_384_XNNPACK_FP32,
         SIZE_384: {
+          DEFAULT: YOLO26_SMALL_384_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_SMALL_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_SMALL_384_COREML_FP16,
         },
         SIZE_512: {
+          DEFAULT: YOLO26_SMALL_512_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_SMALL_512_XNNPACK_FP32,
           COREML_FP16: YOLO26_SMALL_512_COREML_FP16,
         },
         SIZE_640: {
+          DEFAULT: YOLO26_SMALL_640_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_SMALL_640_XNNPACK_FP32,
           COREML_FP16: YOLO26_SMALL_640_COREML_FP16,
         },
@@ -1569,14 +1575,17 @@ export const models = {
       MEDIUM: {
         DEFAULT: YOLO26_MEDIUM_384_XNNPACK_FP32,
         SIZE_384: {
+          DEFAULT: YOLO26_MEDIUM_384_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_MEDIUM_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_MEDIUM_384_COREML_FP16,
         },
         SIZE_512: {
+          DEFAULT: YOLO26_MEDIUM_512_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_MEDIUM_512_XNNPACK_FP32,
           COREML_FP16: YOLO26_MEDIUM_512_COREML_FP16,
         },
         SIZE_640: {
+          DEFAULT: YOLO26_MEDIUM_640_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_MEDIUM_640_XNNPACK_FP32,
           COREML_FP16: YOLO26_MEDIUM_640_COREML_FP16,
         },
@@ -1587,14 +1596,17 @@ export const models = {
       LARGE: {
         DEFAULT: YOLO26_LARGE_384_XNNPACK_FP32,
         SIZE_384: {
+          DEFAULT: YOLO26_LARGE_384_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_LARGE_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_LARGE_384_COREML_FP16,
         },
         SIZE_512: {
+          DEFAULT: YOLO26_LARGE_512_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_LARGE_512_XNNPACK_FP32,
           COREML_FP16: YOLO26_LARGE_512_COREML_FP16,
         },
         SIZE_640: {
+          DEFAULT: YOLO26_LARGE_640_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_LARGE_640_XNNPACK_FP32,
           COREML_FP16: YOLO26_LARGE_640_COREML_FP16,
         },
@@ -1606,14 +1618,17 @@ export const models = {
       XLARGE: {
         DEFAULT: YOLO26_XLARGE_384_XNNPACK_FP32,
         SIZE_384: {
+          DEFAULT: YOLO26_XLARGE_384_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_XLARGE_384_XNNPACK_FP32,
           COREML_FP16: YOLO26_XLARGE_384_COREML_FP16,
         },
         SIZE_512: {
+          DEFAULT: YOLO26_XLARGE_512_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_XLARGE_512_XNNPACK_FP32,
           COREML_FP16: YOLO26_XLARGE_512_COREML_FP16,
         },
         SIZE_640: {
+          DEFAULT: YOLO26_XLARGE_640_XNNPACK_FP32,
           XNNPACK_FP32: YOLO26_XLARGE_640_XNNPACK_FP32,
           COREML_FP16: YOLO26_XLARGE_640_COREML_FP16,
         },
@@ -1642,9 +1657,18 @@ export const models = {
      */
     YOLO26_POSE: {
       DEFAULT: YOLO26_POSE_384_XNNPACK_FP32,
-      SIZE_384: { XNNPACK_FP32: YOLO26_POSE_384_XNNPACK_FP32 },
-      SIZE_512: { XNNPACK_FP32: YOLO26_POSE_512_XNNPACK_FP32 },
-      SIZE_640: { XNNPACK_FP32: YOLO26_POSE_640_XNNPACK_FP32 },
+      SIZE_384: {
+        DEFAULT: YOLO26_POSE_384_XNNPACK_FP32,
+        XNNPACK_FP32: YOLO26_POSE_384_XNNPACK_FP32,
+      },
+      SIZE_512: {
+        DEFAULT: YOLO26_POSE_512_XNNPACK_FP32,
+        XNNPACK_FP32: YOLO26_POSE_512_XNNPACK_FP32,
+      },
+      SIZE_640: {
+        DEFAULT: YOLO26_POSE_640_XNNPACK_FP32,
+        XNNPACK_FP32: YOLO26_POSE_640_XNNPACK_FP32,
+      },
     },
     /**
      * RF-DETR (Roboflow Detection Transformer) pose keypoint detector
@@ -1710,9 +1734,18 @@ export const models = {
        */
       NANO: {
         DEFAULT: YOLO26_NANO_SEG_384_XNNPACK_FP32,
-        SIZE_384: { XNNPACK_FP32: YOLO26_NANO_SEG_384_XNNPACK_FP32 },
-        SIZE_512: { XNNPACK_FP32: YOLO26_NANO_SEG_512_XNNPACK_FP32 },
-        SIZE_640: { XNNPACK_FP32: YOLO26_NANO_SEG_640_XNNPACK_FP32 },
+        SIZE_384: {
+          DEFAULT: YOLO26_NANO_SEG_384_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_NANO_SEG_384_XNNPACK_FP32,
+        },
+        SIZE_512: {
+          DEFAULT: YOLO26_NANO_SEG_512_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_NANO_SEG_512_XNNPACK_FP32,
+        },
+        SIZE_640: {
+          DEFAULT: YOLO26_NANO_SEG_640_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_NANO_SEG_640_XNNPACK_FP32,
+        },
       },
       /**
        * Small scale YOLO26 instance segmentation model. Balanced latency and
@@ -1720,9 +1753,18 @@ export const models = {
        */
       SMALL: {
         DEFAULT: YOLO26_SMALL_SEG_384_XNNPACK_FP32,
-        SIZE_384: { XNNPACK_FP32: YOLO26_SMALL_SEG_384_XNNPACK_FP32 },
-        SIZE_512: { XNNPACK_FP32: YOLO26_SMALL_SEG_512_XNNPACK_FP32 },
-        SIZE_640: { XNNPACK_FP32: YOLO26_SMALL_SEG_640_XNNPACK_FP32 },
+        SIZE_384: {
+          DEFAULT: YOLO26_SMALL_SEG_384_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_SMALL_SEG_384_XNNPACK_FP32,
+        },
+        SIZE_512: {
+          DEFAULT: YOLO26_SMALL_SEG_512_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_SMALL_SEG_512_XNNPACK_FP32,
+        },
+        SIZE_640: {
+          DEFAULT: YOLO26_SMALL_SEG_640_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_SMALL_SEG_640_XNNPACK_FP32,
+        },
       },
       /**
        * Medium scale YOLO26 instance segmentation model. Higher mask boundary
@@ -1730,9 +1772,18 @@ export const models = {
        */
       MEDIUM: {
         DEFAULT: YOLO26_MEDIUM_SEG_384_XNNPACK_FP32,
-        SIZE_384: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_384_XNNPACK_FP32 },
-        SIZE_512: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_512_XNNPACK_FP32 },
-        SIZE_640: { XNNPACK_FP32: YOLO26_MEDIUM_SEG_640_XNNPACK_FP32 },
+        SIZE_384: {
+          DEFAULT: YOLO26_MEDIUM_SEG_384_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_MEDIUM_SEG_384_XNNPACK_FP32,
+        },
+        SIZE_512: {
+          DEFAULT: YOLO26_MEDIUM_SEG_512_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_MEDIUM_SEG_512_XNNPACK_FP32,
+        },
+        SIZE_640: {
+          DEFAULT: YOLO26_MEDIUM_SEG_640_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_MEDIUM_SEG_640_XNNPACK_FP32,
+        },
       },
       /**
        * Large scale YOLO26 instance segmentation model. High accuracy instance
@@ -1740,9 +1791,18 @@ export const models = {
        */
       LARGE: {
         DEFAULT: YOLO26_LARGE_SEG_384_XNNPACK_FP32,
-        SIZE_384: { XNNPACK_FP32: YOLO26_LARGE_SEG_384_XNNPACK_FP32 },
-        SIZE_512: { XNNPACK_FP32: YOLO26_LARGE_SEG_512_XNNPACK_FP32 },
-        SIZE_640: { XNNPACK_FP32: YOLO26_LARGE_SEG_640_XNNPACK_FP32 },
+        SIZE_384: {
+          DEFAULT: YOLO26_LARGE_SEG_384_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_LARGE_SEG_384_XNNPACK_FP32,
+        },
+        SIZE_512: {
+          DEFAULT: YOLO26_LARGE_SEG_512_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_LARGE_SEG_512_XNNPACK_FP32,
+        },
+        SIZE_640: {
+          DEFAULT: YOLO26_LARGE_SEG_640_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_LARGE_SEG_640_XNNPACK_FP32,
+        },
       },
       /**
        * Extra Large scale YOLO26 instance segmentation model. Maximum instance
@@ -1750,9 +1810,18 @@ export const models = {
        */
       XLARGE: {
         DEFAULT: YOLO26_XLARGE_SEG_384_XNNPACK_FP32,
-        SIZE_384: { XNNPACK_FP32: YOLO26_XLARGE_SEG_384_XNNPACK_FP32 },
-        SIZE_512: { XNNPACK_FP32: YOLO26_XLARGE_SEG_512_XNNPACK_FP32 },
-        SIZE_640: { XNNPACK_FP32: YOLO26_XLARGE_SEG_640_XNNPACK_FP32 },
+        SIZE_384: {
+          DEFAULT: YOLO26_XLARGE_SEG_384_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_XLARGE_SEG_384_XNNPACK_FP32,
+        },
+        SIZE_512: {
+          DEFAULT: YOLO26_XLARGE_SEG_512_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_XLARGE_SEG_512_XNNPACK_FP32,
+        },
+        SIZE_640: {
+          DEFAULT: YOLO26_XLARGE_SEG_640_XNNPACK_FP32,
+          XNNPACK_FP32: YOLO26_XLARGE_SEG_640_XNNPACK_FP32,
+        },
       },
     },
   },

From 150a4b5478905ddf8c6d4d1bd2d2b58db5896c11 Mon Sep 17 00:00:00 2001
From: Bartosz Hanc <bartosz.hanc02@gmail.com>
Date: Wed, 26 Aug 2026 15:11:40 +0200
Subject: [PATCH 38/38] docs(runtime): format WorkletRuntime as code in JSDoc
 to avoid external link warning

---
 packages/react-native-executorch/src/core/runtime.ts | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/packages/react-native-executorch/src/core/runtime.ts b/packages/react-native-executorch/src/core/runtime.ts
index 6c68dcc340..9d41837dc9 100644
--- a/packages/react-native-executorch/src/core/runtime.ts
+++ b/packages/react-native-executorch/src/core/runtime.ts
@@ -18,7 +18,7 @@ import { isRnExecuTorchError, RnExecuTorchError } from './error';
  *
  * This runtime runs on a dedicated thread separate from the React Native JS
  * thread, preventing model loading and inference from blocking the UI. Pass it
- * explicitly (or a custom {@link WorkletRuntime}) to {@link wrapAsync} when you
+ * explicitly (or a custom `WorkletRuntime`) to {@link wrapAsync} when you
  * need fine-grained control over which thread work executes on.
  * @category Core / Constants
  */
@@ -28,7 +28,7 @@ export const defaultWorkletRuntime = createWorkletRuntime({
 
 /**
  * Wraps a synchronous worklet function so that it runs asynchronously on a
- * background {@link WorkletRuntime} thread and returns a `Promise`.
+ * background `WorkletRuntime` thread and returns a `Promise`.
  *
  * The wrapper serializes arguments, dispatches the worklet to the target
  * runtime, awaits the result, and re-throws any error thrown inside the worklet