local-agent-memory 1.0.7 → 1.0.8

package/dist/core/search.js

@@ -27,7 +27,7 @@ async function hybridSearch(query, limit, db, filter) {
     const asyncDb = new sync_1.AsyncDatabase(db);
     const RRF_K = 60;
     // 1. FTS5 Search
-    const ftsResults = await new Promise((resolve, reject) => {
+    const ftsResults = await new Promise((resolve) => {
         // Sanitize query to prevent FTS5 syntax errors
         const ftsQuery = sanitizeFtsQuery(query);
         if (!ftsQuery) {
@@ -35,27 +35,71 @@ async function hybridSearch(query, limit, db, filter) {
         }
         // using bm25() for scoring in FTS5
         db.all(`SELECT rowid as id, rank as score FROM chunks_fts WHERE chunks_fts MATCH ? ORDER BY rank LIMIT 100`, [ftsQuery], (err, rows) => {
-            if (err)
-                reject(err);
-            else
-                resolve(rows);
+            if (err) {
+                resolve([]);
+            }
+            else {
+                resolve(rows || []);
+            }
         });
     });
-    // 2. Vector Search
-    const queryEmbedding = await (0, embeddings_1.generateEmbedding)(query);
-    const vectorData = Buffer.from(new Float32Array(queryEmbedding).buffer);
-    const vecResults = await new Promise((resolve, reject) => {
-        // sqlite-vec distance function. distance is lower = better (closer)
-        // we want smallest distance first
-        db.all(`SELECT rowid as id, distance FROM chunks_vec
-       WHERE embedding MATCH ? AND k = 100
-       ORDER BY distance`, [vectorData], (err, rows) => {
-            if (err)
-                reject(err);
-            else
-                resolve(rows);
+    // 2. Vector Search (with graceful fallback)
+    let vecResults = [];
+    try {
+        const queryEmbedding = await (0, embeddings_1.generateEmbedding)(query);
+        const vectorData = Buffer.from(new Float32Array(queryEmbedding).buffer);
+        vecResults = await new Promise((resolve) => {
+            // sqlite-vec distance function. distance is lower = better (closer)
+            // we want smallest distance first
+            db.all(`SELECT rowid as id, distance FROM chunks_vec
+         WHERE embedding MATCH ? AND k = 100
+         ORDER BY distance`, [vectorData], (err, rows) => {
+                if (!err && rows && rows.length > 0) {
+                    return resolve(rows);
+                }
+                // Fallback: If chunks_vec is a regular table (no virtual distance/MATCH)
+                // or sqlite-vec is not loaded, compute cosine similarity in-memory using stored BLOBs.
+                db.all(`SELECT rowid as id, embedding FROM chunks_vec WHERE embedding IS NOT NULL`, [], (fallbackErr, fallbackRows) => {
+                    if (fallbackErr || !fallbackRows || fallbackRows.length === 0) {
+                        return resolve([]);
+                    }
+                    try {
+                        const queryFloats = new Float32Array(queryEmbedding);
+                        const scored = [];
+                        for (const row of fallbackRows) {
+                            if (!row.embedding)
+                                continue;
+                            const buf = row.embedding;
+                            const storedFloats = new Float32Array(buf.buffer, buf.byteOffset, buf.byteLength / 4);
+                            if (storedFloats.length !== queryFloats.length)
+                                continue;
+                            let dotProduct = 0;
+                            for (let i = 0; i < queryFloats.length; i++) {
+                                const q = queryFloats[i];
+                                const s = storedFloats[i];
+                                if (q !== undefined && s !== undefined) {
+                                    dotProduct += q * s;
+                                }
+                            }
+                            // MiniLM embeddings are normalized, so cosine_distance = 1 - dotProduct
+                            scored.push({
+                                id: row.id,
+                                distance: 1 - dotProduct,
+                            });
+                        }
+                        scored.sort((a, b) => a.distance - b.distance);
+                        resolve(scored.slice(0, 100));
+                    }
+                    catch {
+                        resolve([]);
+                    }
+                });
+            });
         });
-    });
+    }
+    catch {
+        vecResults = [];
+    }
     // 3. RRF Fusion
     const rrfScores = new Map();
     // Add FTS scores

package/dist/db/project.d.ts

@@ -4,6 +4,7 @@ export interface FileManifestRow {
     mtime: number;
     sha256: string;
 }
+export declare function getVecExtensionPath(): string | null;
 export declare function getLocalDbPath(projectRoot: string): string;
 export declare function initLocalDb(dbPath: string): sqlite3.Database;
 export declare function upsertFileManifest(dbPath: string, file_path: string, mtime: number, sha256: string): Promise<void>;

package/dist/db/project.js

@@ -3,14 +3,33 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
     return (mod && mod.__esModule) ? mod : { "default": mod };
 };
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.getVecExtensionPath = getVecExtensionPath;
 exports.getLocalDbPath = getLocalDbPath;
 exports.initLocalDb = initLocalDb;
 exports.upsertFileManifest = upsertFileManifest;
 exports.getFileManifest = getFileManifest;
 exports.getAllManifests = getAllManifests;
 exports.removeFileManifest = removeFileManifest;
+const node_fs_1 = __importDefault(require("node:fs"));
+const node_os_1 = __importDefault(require("node:os"));
 const node_path_1 = __importDefault(require("node:path"));
 const sqlite3_1 = __importDefault(require("sqlite3"));
+function getVecExtensionPath() {
+    try {
+        const memDir = process.env.MEMORY_GLOBAL_DIR || node_path_1.default.join(node_os_1.default.homedir(), ".local_memory");
+        const binDir = node_path_1.default.join(memDir, "bin");
+        const extFile = process.platform === "win32"
+            ? "vec0.dll"
+            : process.platform === "darwin"
+                ? "vec0.dylib"
+                : "vec0.so";
+        const extPath = node_path_1.default.join(binDir, extFile);
+        return node_fs_1.default.existsSync(extPath) ? extPath : null;
+    }
+    catch {
+        return null;
+    }
+}
 function getLocalDbPath(projectRoot) {
     return node_path_1.default.join(projectRoot, ".memory", "local.sqlite");
 }
@@ -44,22 +63,41 @@ function initLocalDb(dbPath) {
         endLine INTEGER
       )
     `);
-        // Note: for real usage, chunks_fts and chunks_vec need FTS5 and vec0 extensions enabled.
-        // Standard sqlite3 node driver has FTS5, but vec0 must be loaded.
-        // Assuming they are available or using regular tables for E2E.
         db.run(`
       CREATE VIRTUAL TABLE IF NOT EXISTS chunks_fts USING fts5(
         path, text, breadcrumbs, content='chunks', content_rowid='id'
       )
     `);
-        // Create a regular table for chunks_vec to prevent crashes in tests or environments
-        // where the sqlite-vec extension is not loaded yet.
-        db.run(`
-      CREATE TABLE IF NOT EXISTS chunks_vec (
-        rowid INTEGER,
-        embedding BLOB
-      )
-    `);
+        // Load sqlite-vec extension if downloaded, and create virtual table vec0.
+        // Falls back to regular table if extension is not present or fails to load.
+        const extPath = getVecExtensionPath();
+        if (extPath) {
+            db.loadExtension(extPath, (err) => {
+                if (!err) {
+                    db.run(`
+            CREATE VIRTUAL TABLE IF NOT EXISTS chunks_vec USING vec0(
+              embedding float[384]
+            )
+          `);
+                }
+                else {
+                    db.run(`
+            CREATE TABLE IF NOT EXISTS chunks_vec (
+              rowid INTEGER,
+              embedding BLOB
+            )
+          `);
+                }
+            });
+        }
+        else {
+            db.run(`
+        CREATE TABLE IF NOT EXISTS chunks_vec (
+          rowid INTEGER,
+          embedding BLOB
+        )
+      `);
+        }
     });
     return db;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "local-agent-memory",
-  "version": "1.0.7",
+  "version": "1.0.8",
   "description": "",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",