feat: extend AI tagging to video files via YouTube-style keyframe sampling

backend/app/config/settings.py

@@ -39,6 +39,7 @@
 else:
     DATABASE_PATH = os.path.join(user_data_dir("PictoPy"), "database", "PictoPy.db")
 THUMBNAIL_IMAGES_PATH = os.path.join(user_data_dir("PictoPy"), "thumbnails")
+VIDEO_FRAMES_PATH = os.path.join(user_data_dir("PictoPy"), "video_frames")
 IMAGES_PATH = "./images"
 
 
@@ -175,6 +176,33 @@ def _get_env_int(
 }
 SEMANTIC_DEFAULT_THRESHOLD = 5e-05
 
+# Video keyframe sampling. One frame every N seconds instead of per-frame
+# inference (30fps = 1800 forward passes per minute of video). The interval is
+# stretched when a video would exceed the cap, so cost per video is bounded.
+# The bounds are the single source of truth for both the sampler default here
+# and the user-preferences API schema, so the API can never accept an interval
+# the sampler would reject.
+VIDEO_FRAME_INTERVAL_MIN = 0.5
+VIDEO_FRAME_INTERVAL_MAX = 300.0
+VIDEO_FRAME_INTERVAL_SECONDS = _get_env_float(
+    "VIDEO_FRAME_INTERVAL_SECONDS",
+    5.0,
+    min_value=VIDEO_FRAME_INTERVAL_MIN,
+    max_value=VIDEO_FRAME_INTERVAL_MAX,
+)
+VIDEO_MAX_FRAMES_PER_VIDEO = _get_env_int(
+    "VIDEO_MAX_FRAMES_PER_VIDEO", 200, min_value=1
+)
+# Frames are saved above SigLIP2's largest input (384) so a checkpoint swap
+# doesn't require re-extraction.
+VIDEO_FRAME_MAX_DIMENSION = _get_env_int("VIDEO_FRAME_MAX_DIMENSION", 640, min_value=64)
+# A tag must appear in this many frames to describe the video; drops one-off
+# detections from a single unlucky keyframe.
+VIDEO_TAG_MIN_FRAME_SUPPORT = _get_env_int(
+    "VIDEO_TAG_MIN_FRAME_SUPPORT", 2, min_value=1
+)
+VIDEO_TAG_TOP_K = _get_env_int("VIDEO_TAG_TOP_K", 15, min_value=1)
+
 # Clustering Configuration
 PICTO_CLUSTERING_EPS = _get_env_float("PICTO_CLUSTERING_EPS", 0.75, min_value=0.0)
 PICTO_CLUSTERING_MIN_SAMPLES = _get_env_int(

backend/app/database/folders.py

@@ -408,30 +408,35 @@ def db_get_folder_ids_by_paths(
 
 
 def db_get_all_folder_details() -> (
-    List[Tuple[str, str, Optional[str], int, bool, Optional[bool], str, int]]
+    List[Tuple[str, str, Optional[str], int, bool, Optional[bool], str, int, int]]
 ):
     """
     Get all folder details including folder_id, folder_path, parent_folder_id,
-    last_modified_time, AI_Tagging, taggingCompleted, and image_count.
+    last_modified_time, AI_Tagging, taggingCompleted, indexing_status,
+    image_count and video_count.
     Returns list of tuples with all folder information.
     """
     conn = sqlite3.connect(DATABASE_PATH)
     cursor = conn.cursor()
 
     try:
+        # COUNT(DISTINCT ...) because joining both media tables multiplies the
+        # rows: a folder with 3 images and 4 videos yields 12 join rows.
         cursor.execute(
             """
-            SELECT 
-                f.folder_id, 
-                f.folder_path, 
-                f.parent_folder_id, 
-                f.last_modified_time, 
-                f.AI_Tagging, 
+            SELECT
+                f.folder_id,
+                f.folder_path,
+                f.parent_folder_id,
+                f.last_modified_time,
+                f.AI_Tagging,
                 f.taggingCompleted,
                 f.indexing_status,
-                COUNT(i.id) as image_count
+                COUNT(DISTINCT i.id) as image_count,
+                COUNT(DISTINCT v.id) as video_count
             FROM folders f
             LEFT JOIN images i ON f.folder_id = i.folder_id
+            LEFT JOIN videos v ON f.folder_id = v.folder_id
             GROUP BY f.folder_id
             ORDER BY f.folder_path
             """