feat: multi-provider embeddings, distillation vector search, and cross-project recall

src/config.ts

@@ -66,21 +66,29 @@ export const LoreConfig = z.object({
        *  When enabled, the configured model generates 2–3 alternative query phrasings
        *  before search, improving recall for ambiguous queries. */
       queryExpansion: z.boolean().default(false),
-      /** Vector embedding search via Voyage AI.
-       *  Automatically enabled when VOYAGE_API_KEY env var is set.
+      /** Vector embedding search.
+       *  Supports multiple providers: "voyage" (Voyage AI, VOYAGE_API_KEY),
+       *  "openai" (OpenAI, OPENAI_API_KEY).
+       *  Automatically enabled when the configured provider's API key env var is set.
        *  Set enabled: false to explicitly disable even with the key present. */
       embeddings: z
         .object({
           /** Enable/disable vector embedding search. Default: true.
-           *  Set to false to explicitly disable even when VOYAGE_API_KEY is set. */
+           *  Set to false to explicitly disable even when the API key is set. */
           enabled: z.boolean().default(true),
-          /** Voyage AI model ID. Default: voyage-code-3. */
+          /** Embedding provider. Default: "voyage".
+           *  Each provider reads its own env var for the API key:
+           *  - "voyage": VOYAGE_API_KEY (default model: voyage-code-3, 1024 dims)
+           *  - "openai": OPENAI_API_KEY (default model: text-embedding-3-small, 1536 dims) */
+          provider: z.enum(["voyage", "openai"]).default("voyage"),
+          /** Model ID for the embedding provider. Default depends on provider. */
           model: z.string().default("voyage-code-3"),
           /** Embedding dimensions. Default: 1024. */
           dimensions: z.number().min(256).max(2048).default(1024),
         })
         .default({
           enabled: true,
+          provider: "voyage",
           model: "voyage-code-3",
           dimensions: 1024,
         }),
@@ -89,7 +97,7 @@ export const LoreConfig = z.object({
       ftsWeights: { title: 6.0, content: 2.0, category: 3.0 },
       recallLimit: 10,
       queryExpansion: false,
-      embeddings: { enabled: true, model: "voyage-code-3", dimensions: 1024 },
+      embeddings: { enabled: true, provider: "voyage" as const, model: "voyage-code-3", dimensions: 1024 },
     }),
   crossProject: z.boolean().default(false),
   agentsFile: z

src/db.ts

@@ -2,7 +2,7 @@ import { Database } from "bun:sqlite";
 import { join, dirname } from "path";
 import { mkdirSync } from "fs";
 
-const SCHEMA_VERSION = 8;
+const SCHEMA_VERSION = 9;
 
 const MIGRATIONS: string[] = [
   `
@@ -220,6 +220,14 @@ const MIGRATIONS: string[] = [
     value TEXT NOT NULL
   );
   `,
+  `
+  -- Version 9: Embedding BLOB column for distillation vector search.
+  -- Same pattern as knowledge embeddings (version 8). Enables semantic
+  -- search over distilled session summaries via cosine similarity.
+  -- No backfill — entries get embedded lazily on next distillation
+  -- or via explicit backfill when embeddings are first enabled.
+  ALTER TABLE distillations ADD COLUMN embedding BLOB;
+  `,
 ];
 
 function dataDir() {
@@ -316,6 +324,14 @@ export function projectId(path: string): string | undefined {
   return row?.id;
 }
 
+/** Look up a project's display name by its internal ID. */
+export function projectName(id: string): string | null {
+  const row = db()
+    .query("SELECT name FROM projects WHERE id = ?")
+    .get(id) as { name: string } | null;
+  return row?.name ?? null;
+}
+
 /**
  * Returns true if Lore has never been used before (no projects in the DB).
  * Must be called before ensureProject() to get an accurate result.

src/distillation.ts

@@ -2,6 +2,7 @@ import type { createOpencodeClient } from "@opencode-ai/sdk";
 import { db, ensureProject } from "./db";
 import { config } from "./config";
 import * as temporal from "./temporal";
+import * as embedding from "./embedding";
 import * as log from "./log";
 import {
   DISTILLATION_SYSTEM,
@@ -402,14 +403,20 @@ async function distillSegment(input: {
   const result = parseDistillationResult(responsePart.text);
   if (!result) return null;
 
-  storeDistillation({
+  const distillId = storeDistillation({
     projectPath: input.projectPath,
     sessionID: input.sessionID,
     observations: result.observations,
     sourceIDs: input.messages.map((m) => m.id),
     generation: 0,
   });
   temporal.markDistilled(input.messages.map((m) => m.id));
+
+  // Fire-and-forget: embed the distillation for vector search
+  if (embedding.isAvailable()) {
+    embedding.embedDistillation(distillId, result.observations);
+  }
+
   return result;
 }
 
@@ -458,14 +465,19 @@ async function metaDistill(input: {
   // Store the meta-distillation at generation N+1
   const maxGen = Math.max(...existing.map((d) => d.generation));
   const allSourceIDs = existing.flatMap((d) => d.source_ids);
-  storeDistillation({
+  const metaId = storeDistillation({
     projectPath: input.projectPath,
     sessionID: input.sessionID,
     observations: result.observations,
     sourceIDs: allSourceIDs,
     generation: maxGen + 1,
   });
 
+  // Fire-and-forget: embed the meta-distillation for vector search
+  if (embedding.isAvailable()) {
+    embedding.embedDistillation(metaId, result.observations);
+  }
+
   // Archive the gen-0 distillations that were merged into gen-1+.
   // They remain searchable via recall but excluded from the in-context prefix.
   archiveDistillations(existing.map((d) => d.id));