ruvector 0.2.26 → 0.2.27

package/dist/core/adaptive-embedder.d.ts

@@ -81,6 +81,15 @@ export declare class AdaptiveEmbedder {
      * Initialize ONNX backend
      */
     init(): Promise<void>;
+    /**
+     * Whether the ONNX backend is initialized and producing real (non-hash)
+     * embeddings. Returns `false` until `init()` has successfully loaded the
+     * ONNX model. Note that `embed()` still works when this is `false` — it
+     * falls back to a deterministic hash embedding — so callers that require
+     * real semantic embeddings should check this after `init()`.
+     * See https://github.com/ruvnet/RuVector/issues/523.
+     */
+    isReady(): boolean;
     /**
      * Generate adaptive embedding
      * Pipeline: ONNX → LoRA → Prototype Adjustment → Episodic Augmentation

package/dist/core/adaptive-embedder.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"adaptive-embedder.d.ts","sourceRoot":"","sources":["../../src/core/adaptive-embedder.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AAQH,MAAM,WAAW,cAAc;IAC7B,iEAAiE;IACjE,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,uCAAuC;IACvC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,kCAAkC;IAClC,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,8CAA8C;IAC9C,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,gDAAgD;IAChD,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,uCAAuC;IACvC,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,6CAA6C;IAC7C,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,WAAW;IAC1B,CAAC,EAAE,MAAM,EAAE,EAAE,CAAC;IACd,CAAC,EAAE,MAAM,EAAE,EAAE,CAAC;IACd,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;CACjB;AAED,MAAM,WAAW,eAAe;IAC9B,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,aAAa;IAC5B,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,kBAAkB,EAAE,MAAM,CAAC;CAC5B;AA+pBD,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,MAAM,CAA2B;IACzC,OAAO,CAAC,IAAI,CAAY;IACxB,OAAO,CAAC,UAAU,CAAkB;IACpC,OAAO,CAAC,QAAQ,CAAiB;IACjC,OAAO,CAAC,SAAS,CAAkB;IACnC,OAAO,CAAC,SAAS,CAAe;IAGhC,OAAO,CAAC,eAAe,CAAa;IACpC,OAAO,CAAC,QAAQ,CAAa;IAC7B,OAAO,CAAC,gBAAgB,CAAa;IAGrC,OAAO,CAAC,YAAY,CAA+E;gBAEvF,MAAM,GAAE,cAAmB;IAiBvC;;OAEG;IACG,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAO3B;;;OAGG;IACG,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE;QAClC,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,WAAW,CAAC,EAAE,OAAO,CAAC;QACtB,aAAa,CAAC,EAAE,OAAO,CAAC;KACzB,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;IAmCrB;;OAEG;IACG,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,EAAE,OAAO,CAAC,EAAE;QAC1C,MAAM,CAAC,EAAE,MAAM,CAAC;KACjB,GAAG,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC;IAsBvB;;;OAGG;IACG,WAAW,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAkBpG;;OAEG;IACH,OAAO,CAAC,kBAAkB;IA+B1B;;OAEG;IACG,gBAAgB,CACpB,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,OAAO,EAChB,OAAO,GAAE,MAAY,GACpB,OAAO,CAAC,IAAI,CAAC;IAiBhB;;;OAGG;IACG,WAAW,IAAI,OAAO,CAAC,IAAI,CAAC;IAmBlC;;OAEG;IACH,OAAO,CAAC,SAAS;IAoBjB,OAAO,CAAC,SAAS;IAKjB;;OAEG;IACH,QAAQ,IAAI,aAAa;IAczB;;OAEG;IACH,MAAM,IAAI;QACR,IAAI,EAAE,WAAW,CAAC;QAClB,UAAU,EAAE,eAAe,EAAE,CAAC;QAC9B,KAAK,EAAE,aAAa,CAAC;KACtB;IAQD;;OAEG;IACH,MAAM,CAAC,IAAI,EAAE;QAAE,IAAI,CAAC,EAAE,WAAW,CAAC;QAAC,UAAU,CAAC,EAAE,eAAe,EAAE,CAAA;KAAE,GAAG,IAAI;IAS1E;;OAEG;IACH,KAAK,IAAI,IAAI;IAUb;;OAEG;IACH,aAAa,IAAI;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE;CAGnD;AAQD,wBAAgB,mBAAmB,CAAC,MAAM,CAAC,EAAE,cAAc,GAAG,gBAAgB,CAK7E;AAED,wBAAsB,oBAAoB,CAAC,MAAM,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAI7F;AAED,eAAe,gBAAgB,CAAC"}
+{"version":3,"file":"adaptive-embedder.d.ts","sourceRoot":"","sources":["../../src/core/adaptive-embedder.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AAQH,MAAM,WAAW,cAAc;IAC7B,iEAAiE;IACjE,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,uCAAuC;IACvC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,kCAAkC;IAClC,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,8CAA8C;IAC9C,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,gDAAgD;IAChD,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,uCAAuC;IACvC,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,6CAA6C;IAC7C,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AAED,MAAM,WAAW,WAAW;IAC1B,CAAC,EAAE,MAAM,EAAE,EAAE,CAAC;IACd,CAAC,EAAE,MAAM,EAAE,EAAE,CAAC;IACd,IAAI,CAAC,EAAE,MAAM,EAAE,CAAC;CACjB;AAED,MAAM,WAAW,eAAe;IAC9B,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,MAAM,EAAE,CAAC;IACnB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,aAAa;IAC5B,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,kBAAkB,EAAE,MAAM,CAAC;CAC5B;AA+pBD,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,MAAM,CAA2B;IACzC,OAAO,CAAC,IAAI,CAAY;IACxB,OAAO,CAAC,UAAU,CAAkB;IACpC,OAAO,CAAC,QAAQ,CAAiB;IACjC,OAAO,CAAC,SAAS,CAAkB;IACnC,OAAO,CAAC,SAAS,CAAe;IAGhC,OAAO,CAAC,eAAe,CAAa;IACpC,OAAO,CAAC,QAAQ,CAAa;IAC7B,OAAO,CAAC,gBAAgB,CAAa;IAGrC,OAAO,CAAC,YAAY,CAA+E;gBAEvF,MAAM,GAAE,cAAmB;IAiBvC;;OAEG;IACG,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;IAO3B;;;;;;;OAOG;IACH,OAAO,IAAI,OAAO;IAIlB;;;OAGG;IACG,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE;QAClC,MAAM,CAAC,EAAE,MAAM,CAAC;QAChB,WAAW,CAAC,EAAE,OAAO,CAAC;QACtB,aAAa,CAAC,EAAE,OAAO,CAAC;KACzB,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;IAmCrB;;OAEG;IACG,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,EAAE,OAAO,CAAC,EAAE;QAC1C,MAAM,CAAC,EAAE,MAAM,CAAC;KACjB,GAAG,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC;IAsBvB;;;OAGG;IACG,WAAW,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAkBpG;;OAEG;IACH,OAAO,CAAC,kBAAkB;IA+B1B;;OAEG;IACG,gBAAgB,CACpB,OAAO,EAAE,MAAM,EACf,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,OAAO,EAChB,OAAO,GAAE,MAAY,GACpB,OAAO,CAAC,IAAI,CAAC;IAiBhB;;;OAGG;IACG,WAAW,IAAI,OAAO,CAAC,IAAI,CAAC;IAmBlC;;OAEG;IACH,OAAO,CAAC,SAAS;IAoBjB,OAAO,CAAC,SAAS;IAKjB;;OAEG;IACH,QAAQ,IAAI,aAAa;IAczB;;OAEG;IACH,MAAM,IAAI;QACR,IAAI,EAAE,WAAW,CAAC;QAClB,UAAU,EAAE,eAAe,EAAE,CAAC;QAC9B,KAAK,EAAE,aAAa,CAAC;KACtB;IAQD;;OAEG;IACH,MAAM,CAAC,IAAI,EAAE;QAAE,IAAI,CAAC,EAAE,WAAW,CAAC;QAAC,UAAU,CAAC,EAAE,eAAe,EAAE,CAAA;KAAE,GAAG,IAAI;IAS1E;;OAEG;IACH,KAAK,IAAI,IAAI;IAUb;;OAEG;IACH,aAAa,IAAI;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE;CAGnD;AAQD,wBAAgB,mBAAmB,CAAC,MAAM,CAAC,EAAE,cAAc,GAAG,gBAAgB,CAK7E;AAED,wBAAsB,oBAAoB,CAAC,MAAM,CAAC,EAAE,cAAc,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAI7F;AAED,eAAe,gBAAgB,CAAC"}

package/dist/core/adaptive-embedder.js

@@ -611,6 +611,17 @@ class AdaptiveEmbedder {
             this.onnxReady = true;
         }
     }
+    /**
+     * Whether the ONNX backend is initialized and producing real (non-hash)
+     * embeddings. Returns `false` until `init()` has successfully loaded the
+     * ONNX model. Note that `embed()` still works when this is `false` — it
+     * falls back to a deterministic hash embedding — so callers that require
+     * real semantic embeddings should check this after `init()`.
+     * See https://github.com/ruvnet/RuVector/issues/523.
+     */
+    isReady() {
+        return this.onnxReady;
+    }
     /**
      * Generate adaptive embedding
      * Pipeline: ONNX → LoRA → Prototype Adjustment → Episodic Augmentation

package/dist/core/onnx/bundled-parallel.mjs

@@ -0,0 +1,164 @@
+/**
+ * Bundled-WASM parallel embedder (issue #523 SOTA).
+ *
+ * A self-contained worker_threads pool — NO external dependency — that shards
+ * batches of text across CPU cores, each worker running the bundled ONNX WASM
+ * embedder over the SAME model bytes (shared via SharedArrayBuffer) and config.
+ * Output vectors are identical to the single-thread path (cosine-equivalent),
+ * so this is a pure throughput optimization with no quality change.
+ *
+ * Drop-in shape compatible with the optional `ruvector-onnx-embeddings-wasm/parallel`
+ * package: { numWorkers, dimension, init(), embedBatch(texts) -> number[][], shutdown() }.
+ */
+import { Worker } from 'node:worker_threads';
+import * as os from 'node:os';
+import { fileURLToPath } from 'node:url';
+import * as path from 'node:path';
+
+const __dirname = path.dirname(fileURLToPath(import.meta.url));
+
+export class ParallelEmbedder {
+  /**
+   * @param {object} opts
+   * @param {Uint8Array} opts.modelBytes  raw ONNX model bytes (loaded once by caller)
+   * @param {string}     opts.tokenizerJson
+   * @param {number}     [opts.maxLength=256]
+   * @param {number}     [opts.dimension=384]
+   * @param {number}     [opts.numWorkers]  defaults to min(cpus-2, 16), >=2
+   */
+  constructor(opts = {}) {
+    this.numWorkers = opts.numWorkers || Math.max(2, Math.min((os.cpus().length || 4) - 2, 16));
+    this.dimension = opts.dimension || 384;
+    this._modelBytes = opts.modelBytes;
+    this._tokenizerJson = opts.tokenizerJson;
+    this._maxLength = opts.maxLength || 256;
+    this._requestTimeoutMs = opts.requestTimeoutMs ?? 30000;
+    this._workers = [];
+    this._pending = new Map();   // id -> { resolve, reject, worker, timer }
+    this._seq = 0;
+    this._shuttingDown = false;
+  }
+
+  async init() {
+    if (!this._modelBytes || !this._tokenizerJson) {
+      throw new Error('ParallelEmbedder requires modelBytes and tokenizerJson');
+    }
+    // Share model bytes across all workers via a single SharedArrayBuffer.
+    const sab = new SharedArrayBuffer(this._modelBytes.length);
+    new Uint8Array(sab).set(this._modelBytes);
+
+    const workerUrl = new URL('./embed-worker.mjs', import.meta.url);
+    const readies = [];
+
+    for (let i = 0; i < this.numWorkers; i++) {
+      const w = new Worker(workerUrl, {
+        workerData: { modelSab: sab, tokenizerJson: this._tokenizerJson, maxLength: this._maxLength },
+      });
+      w.on('message', (m) => this._onMessage(m));
+      // If a worker dies (uncaught error or unexpected exit), fail every request
+      // currently routed to it instead of letting those promises hang forever.
+      w.on('error', (e) => this._failWorker(w, e instanceof Error ? e : new Error(String(e))));
+      w.on('exit', (code) => {
+        if (!this._shuttingDown && code !== 0) {
+          this._failWorker(w, new Error(`embed worker exited unexpectedly (code ${code})`));
+        }
+      });
+      this._workers.push(w);
+      readies.push(new Promise((resolve, reject) => {
+        const onReady = (m) => {
+          if (m.type === 'ready') { cleanup(); resolve(); }
+          else if (m.type === 'init-error') { cleanup(); reject(new Error('worker init failed: ' + m.error)); }
+        };
+        const onErr = (e) => { cleanup(); reject(e); };
+        const cleanup = () => { w.off('message', onReady); w.off('error', onErr); };
+        w.on('message', onReady);
+        w.once('error', onErr);
+      }));
+    }
+
+    await Promise.all(readies);
+    // Drop the main-thread reference; the SAB keeps the shared copy alive.
+    this._modelBytes = null;
+  }
+
+  _settle(id, fn) {
+    const p = this._pending.get(id);
+    if (!p) return;
+    this._pending.delete(id);
+    if (p.timer) clearTimeout(p.timer);
+    fn(p);
+  }
+
+  _onMessage(m) {
+    if (m.type !== 'result' && m.type !== 'error') return;
+    this._settle(m.id, (p) => {
+      if (m.type === 'error') p.reject(new Error(m.error));
+      else p.resolve({ dim: m.dim, count: m.count, flat: new Float32Array(m.buffer) });
+    });
+  }
+
+  /** Reject every in-flight request routed to a dead worker. */
+  _failWorker(worker, err) {
+    for (const [id, p] of this._pending) {
+      if (p.worker === worker) this._settle(id, () => p.reject(err));
+    }
+  }
+
+  _send(worker, texts) {
+    const id = ++this._seq;
+    return new Promise((resolve, reject) => {
+      const entry = { resolve, reject, worker, timer: null };
+      if (this._requestTimeoutMs > 0) {
+        entry.timer = setTimeout(() => {
+          this._settle(id, () =>
+            reject(new Error(`embed request timed out after ${this._requestTimeoutMs}ms`)));
+        }, this._requestTimeoutMs);
+        // Don't keep the event loop alive solely for this timer.
+        if (typeof entry.timer.unref === 'function') entry.timer.unref();
+      }
+      this._pending.set(id, entry);
+      worker.postMessage({ type: 'embed', id, texts });
+    });
+  }
+
+  /**
+   * Embed many texts, sharded across workers. Returns number[][] in input order.
+   */
+  async embedBatch(texts) {
+    if (!texts || texts.length === 0) return [];
+    const n = this._workers.length;
+    const shard = Math.ceil(texts.length / n);
+    const tasks = [];
+    const starts = [];
+    for (let i = 0; i < n; i++) {
+      const start = i * shard;
+      if (start >= texts.length) break;
+      const end = Math.min(texts.length, start + shard);
+      starts.push(start);
+      tasks.push(this._send(this._workers[i], texts.slice(start, end)));
+    }
+    const results = await Promise.all(tasks);
+    const out = new Array(texts.length);
+    for (let r = 0; r < results.length; r++) {
+      const { dim, count, flat } = results[r];
+      const start = starts[r];
+      for (let j = 0; j < count; j++) {
+        out[start + j] = Array.from(flat.subarray(j * dim, (j + 1) * dim));
+      }
+    }
+    return out;
+  }
+
+  async shutdown() {
+    this._shuttingDown = true;
+    // Reject anything still in flight so callers don't hang on shutdown.
+    for (const [id, p] of this._pending) {
+      this._settle(id, () => p.reject(new Error('ParallelEmbedder shut down')));
+    }
+    const ws = this._workers;
+    this._workers = [];
+    await Promise.all(ws.map((w) => w.terminate()));
+  }
+}
+
+export default ParallelEmbedder;

package/dist/core/onnx/embed-worker.mjs

@@ -0,0 +1,67 @@
+/**
+ * Worker-thread entry for the bundled-WASM parallel embedder (issue #523 SOTA).
+ *
+ * Each worker loads its own instance of the bundled ONNX WASM embedder from the
+ * SAME model bytes (shared via SharedArrayBuffer — no per-worker download) and
+ * the SAME config, so the vectors it produces are identical to the single-thread
+ * path (cosine-equivalent by construction).
+ *
+ * Protocol:
+ *   workerData: { modelSab: SharedArrayBuffer, tokenizerJson: string, maxLength: number }
+ *   → posts { type: 'ready' } once the WASM embedder is constructed
+ *   message { type: 'embed', id, texts: string[] }
+ *   → posts { type: 'result', id, dim, count, buffer } (Float32Array buffer, transferred)
+ *   errors → { type: 'error', id, error }
+ */
+import { parentPort, workerData } from 'node:worker_threads';
+import { pathToFileURL, fileURLToPath } from 'node:url';
+import * as path from 'node:path';
+import * as fs from 'node:fs';
+
+const __dirname = path.dirname(fileURLToPath(import.meta.url));
+
+let embedder = null;
+
+async function init() {
+  const bgJsPath = path.join(__dirname, 'pkg', 'ruvector_onnx_embeddings_wasm_bg.js');
+  const wasmPath = path.join(__dirname, 'pkg', 'ruvector_onnx_embeddings_wasm_bg.wasm');
+
+  const wasmModule = await import(pathToFileURL(bgJsPath).href);
+  const wasmBytes = fs.readFileSync(wasmPath);
+  const wasmResult = await WebAssembly.instantiate(wasmBytes, {
+    './ruvector_onnx_embeddings_wasm_bg.js': wasmModule,
+  });
+  const wasmExports = wasmResult.instance.exports;
+  if (typeof wasmModule.__wbg_set_wasm === 'function') wasmModule.__wbg_set_wasm(wasmExports);
+  if (typeof wasmExports.__wbindgen_start === 'function') wasmExports.__wbindgen_start();
+
+  // Reconstruct model bytes from the shared buffer (zero-copy view, then handed
+  // to wasm-bindgen which copies into WASM linear memory).
+  const modelBytes = new Uint8Array(workerData.modelSab);
+
+  const cfg = new wasmModule.WasmEmbedderConfig()
+    .setMaxLength(workerData.maxLength || 256)
+    .setNormalize(true)
+    .setPooling(0); // Mean pooling — matches the single-thread path.
+
+  embedder = wasmModule.WasmEmbedder.withConfig(modelBytes, workerData.tokenizerJson, cfg);
+}
+
+parentPort.on('message', (msg) => {
+  if (msg.type !== 'embed') return;
+  try {
+    const dim = embedder.dimension();
+    const flat = embedder.embedBatch(msg.texts); // length = texts.length * dim
+    const arr = Float32Array.from(flat);
+    parentPort.postMessage(
+      { type: 'result', id: msg.id, dim, count: msg.texts.length, buffer: arr.buffer },
+      [arr.buffer],
+    );
+  } catch (e) {
+    parentPort.postMessage({ type: 'error', id: msg.id, error: e?.message || String(e) });
+  }
+});
+
+init()
+  .then(() => parentPort.postMessage({ type: 'ready' }))
+  .catch((e) => parentPort.postMessage({ type: 'init-error', error: e?.message || String(e) }));

package/dist/core/onnx/loader.js

@@ -76,6 +76,14 @@ export const MODELS = {
  */
 export const DEFAULT_MODEL = 'all-MiniLM-L6-v2';
 
+/**
+ * In-memory memo of loaded models, keyed by model name. Deduplicates the
+ * (re-)download + decode when multiple embedder instances load the same model
+ * in one process. In Node there is no Cache API, so without this every
+ * ModelLoader.loadModel() re-fetches the model from HuggingFace (issue #523).
+ */
+const _inMemoryModelCache = new Map();
+
 /**
  * Model loader with caching support
  */
@@ -97,6 +105,24 @@ export class ModelLoader {
             throw new Error(`Unknown model: ${modelName}. Available: ${Object.keys(MODELS).join(', ')}`);
         }
 
+        // In-memory memo: a second load of the same model in this process reuses
+        // the already-downloaded bytes instead of re-fetching (issue #523).
+        if (this.cache && _inMemoryModelCache.has(modelName)) {
+            return _inMemoryModelCache.get(modelName);
+        }
+
+        // On-disk cache (Node only): models persist across processes so they are
+        // downloaded once, not every run. The browser has the Cache API instead
+        // (handled in fetchWithCache). See issue #523.
+        if (this.cache) {
+            const disk = await this._loadFromDisk(modelName);
+            if (disk) {
+                const cached = { ...disk, config: modelConfig };
+                _inMemoryModelCache.set(modelName, cached);
+                return cached;
+            }
+        }
+
         console.log(`Loading model: ${modelConfig.name} (${modelConfig.size})`);
 
         const [modelBytes, tokenizerJson] = await Promise.all([
@@ -104,11 +130,71 @@ export class ModelLoader {
             this.fetchWithCache(modelConfig.tokenizer, `${modelName}-tokenizer.json`, 'text'),
         ]);
 
-        return {
+        const result = {
             modelBytes: new Uint8Array(modelBytes),
             tokenizerJson,
             config: modelConfig,
         };
+
+        if (this.cache) {
+            _inMemoryModelCache.set(modelName, result);
+            await this._saveToDisk(modelName, result.modelBytes, tokenizerJson);
+        }
+
+        return result;
+    }
+
+    /**
+     * Resolve the Node on-disk cache dir for a model (null in non-Node envs).
+     * Uses dynamic import so this module stays loadable in browsers/bundlers.
+     */
+    async _diskCacheDir(modelName) {
+        if (typeof process === 'undefined' || !process.versions?.node) return null;
+        const home = process.env.RUVECTOR_CACHE_DIR
+            || process.env.HOME || process.env.USERPROFILE || '/tmp';
+        const path = await import('node:path');
+        return path.join(home, '.ruvector', 'models', modelName);
+    }
+
+    /** Load model bytes + tokenizer from the Node disk cache, or null if absent. */
+    async _loadFromDisk(modelName) {
+        const dir = await this._diskCacheDir(modelName);
+        if (!dir) return null;
+        try {
+            const fs = await import('node:fs');
+            const path = await import('node:path');
+            const modelPath = path.join(dir, 'model.onnx');
+            const tokPath = path.join(dir, 'tokenizer.json');
+            if (!fs.existsSync(modelPath) || !fs.existsSync(tokPath)) return null;
+            const modelBytes = new Uint8Array(fs.readFileSync(modelPath));
+            const tokenizerJson = fs.readFileSync(tokPath, 'utf8');
+            if (modelBytes.length === 0 || tokenizerJson.length === 0) return null;
+            console.log(`  Disk cache hit: ${modelName}`);
+            return { modelBytes, tokenizerJson };
+        } catch {
+            return null;
+        }
+    }
+
+    /** Persist model bytes + tokenizer to the Node disk cache (best-effort). */
+    async _saveToDisk(modelName, modelBytes, tokenizerJson) {
+        const dir = await this._diskCacheDir(modelName);
+        if (!dir) return;
+        try {
+            const fs = await import('node:fs');
+            const path = await import('node:path');
+            fs.mkdirSync(dir, { recursive: true });
+            // Write to temp files then rename, so a crash mid-write can't leave a
+            // truncated cache entry that later reads would trust.
+            const mTmp = path.join(dir, 'model.onnx.tmp');
+            const tTmp = path.join(dir, 'tokenizer.json.tmp');
+            fs.writeFileSync(mTmp, Buffer.from(modelBytes));
+            fs.writeFileSync(tTmp, tokenizerJson);
+            fs.renameSync(mTmp, path.join(dir, 'model.onnx'));
+            fs.renameSync(tTmp, path.join(dir, 'tokenizer.json'));
+        } catch {
+            // Cache write is best-effort; embedding still works without it.
+        }
     }
 
     /**

package/dist/core/onnx/package.json

@@ -0,0 +1,3 @@
+{
+  "type": "module"
+}

package/dist/core/onnx-embedder.d.ts

@@ -43,7 +43,13 @@ export interface SimilarityResult {
     timeMs: number;
 }
 /**
- * Check if ONNX embedder is available (bundled files exist)
+ * Check if the ONNX embedder is *available* — i.e. the bundled WASM files are
+ * present and the embedder can be initialized.
+ *
+ * NOTE: This is a capability check, NOT a readiness check. It returns `true`
+ * before `initOnnxEmbedder()` has run (so callers can decide whether to init).
+ * To check whether the model has actually been loaded, use `isOnnxInitialized()`
+ * or `isReady()`. See https://github.com/ruvnet/RuVector/issues/523.
  */
 export declare function isOnnxAvailable(): boolean;
 /**
@@ -72,9 +78,22 @@ export declare function cosineSimilarity(a: number[], b: number[]): number;
  */
 export declare function getDimension(): number;
 /**
- * Check if embedder is ready
+ * Check if the embedder has been initialized (model loaded) and is ready to
+ * embed. Returns `false` until `initOnnxEmbedder()` (or the first `embed()`,
+ * which auto-initializes) has completed successfully.
  */
 export declare function isReady(): boolean;
+/**
+ * Whether the ONNX embedder has been initialized (model loaded).
+ *
+ * Post-init counterpart to `isOnnxAvailable()` (which only checks that the
+ * bundled files exist). Named distinctly from the WASM-core `isInitialized()`
+ * export to avoid a barrel name collision. Equivalent to `isReady()`; provided
+ * as a self-documenting gate so callers can distinguish "bundled" (available)
+ * from "loaded" (initialized). See
+ * https://github.com/ruvnet/RuVector/issues/523.
+ */
+export declare function isOnnxInitialized(): boolean;
 /**
  * Get embedder stats including SIMD and parallel capabilities
  */
@@ -91,6 +110,24 @@ export declare function getStats(): {
  * Shutdown parallel workers (call on exit)
  */
 export declare function shutdown(): Promise<void>;
+/**
+ * Initialize the bundled-WASM worker pool for high-throughput batch embedding
+ * (issue #523 SOTA). Self-contained — uses Node worker_threads + the bundled
+ * WASM over SharedArrayBuffer model bytes, no external dependency. Vectors are
+ * identical to the single-thread path (cosine-equivalent).
+ *
+ * @param numWorkers number of worker threads (default: min(cpus-2, 16))
+ */
+export declare function initParallelEmbedder(numWorkers?: number): Promise<boolean>;
+/**
+ * Batch-embed via the bundled worker pool, sharded across CPU cores. Lazily
+ * starts the pool on first use. Returns embeddings in input order.
+ */
+export declare function embedBatchParallel(texts: string[]): Promise<number[][]>;
+/** Number of active pool workers (0 if the pool isn't started). */
+export declare function getParallelWorkerCount(): number;
+/** Shut down the bundled worker pool and release its threads. */
+export declare function shutdownParallelEmbedder(): Promise<void>;
 export declare class OnnxEmbedder {
     private config;
     constructor(config?: OnnxEmbedderConfig);

package/dist/core/onnx-embedder.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"onnx-embedder.d.ts","sourceRoot":"","sources":["../../src/core/onnx-embedder.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAQH,OAAO,CAAC,MAAM,CAAC;IAEb,IAAI,kBAAkB,EAAE,WAAW,GAAG,SAAS,CAAC;CACjD;AAuBD,MAAM,WAAW,kBAAkB;IACjC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,OAAO,CAAC;IACpB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;;OAKG;IACH,cAAc,CAAC,EAAE,OAAO,GAAG,MAAM,CAAC;IAClC,wDAAwD;IACxD,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,iEAAiE;IACjE,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC5B;AAMD,MAAM,WAAW,eAAe;IAC9B,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,gBAAgB;IAC/B,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CAChB;AAeD;;GAEG;AACH,wBAAgB,eAAe,IAAI,OAAO,CAOzC;AA6DD;;GAEG;AACH,wBAAsB,gBAAgB,CAAC,MAAM,GAAE,kBAAuB,GAAG,OAAO,CAAC,OAAO,CAAC,CAkGxF;AAED;;GAEG;AACH,wBAAsB,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,eAAe,CAAC,CAiBlE;AAED;;;GAGG;AACH,wBAAsB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC,CAwC5E;AAED;;GAEG;AACH,wBAAsB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAaxF;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,MAAM,EAAE,EAAE,CAAC,EAAE,MAAM,EAAE,GAAG,MAAM,CAiBjE;AAED;;GAEG;AACH,wBAAgB,YAAY,IAAI,MAAM,CAErC;AAED;;GAEG;AACH,wBAAgB,OAAO,IAAI,OAAO,CAEjC;AAED;;GAEG;AACH,wBAAgB,QAAQ,IAAI;IAC1B,KAAK,EAAE,OAAO,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,OAAO,CAAC;IACd,QAAQ,EAAE,OAAO,CAAC;IAClB,eAAe,EAAE,MAAM,CAAC;IACxB,iBAAiB,EAAE,MAAM,CAAC;CAC3B,CAUA;AAED;;GAEG;AACH,wBAAsB,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC,CAM9C;AAGD,qBAAa,YAAY;IACvB,OAAO,CAAC,MAAM,CAAqB;gBAEvB,MAAM,GAAE,kBAAuB;IAIrC,IAAI,IAAI,OAAO,CAAC,OAAO,CAAC;IAIxB,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;IAKtC,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC;IAKhD,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAK/D,IAAI,SAAS,IAAI,MAAM,CAEtB;IAED,IAAI,KAAK,IAAI,OAAO,CAEnB;CACF;AAED,eAAe,YAAY,CAAC"}
+{"version":3,"file":"onnx-embedder.d.ts","sourceRoot":"","sources":["../../src/core/onnx-embedder.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAQH,OAAO,CAAC,MAAM,CAAC;IAEb,IAAI,kBAAkB,EAAE,WAAW,GAAG,SAAS,CAAC;CACjD;AAuBD,MAAM,WAAW,kBAAkB;IACjC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,OAAO,CAAC;IACpB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB;;;;;OAKG;IACH,cAAc,CAAC,EAAE,OAAO,GAAG,MAAM,CAAC;IAClC,wDAAwD;IACxD,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,iEAAiE;IACjE,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC5B;AAMD,MAAM,WAAW,eAAe;IAC9B,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,gBAAgB;IAC/B,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;CAChB;AAsBD;;;;;;;;GAQG;AACH,wBAAgB,eAAe,IAAI,OAAO,CAOzC;AA0FD;;GAEG;AACH,wBAAsB,gBAAgB,CAAC,MAAM,GAAE,kBAAuB,GAAG,OAAO,CAAC,OAAO,CAAC,CAuGxF;AAED;;GAEG;AACH,wBAAsB,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,eAAe,CAAC,CAiBlE;AAED;;;GAGG;AACH,wBAAsB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC,CAwC5E;AAED;;GAEG;AACH,wBAAsB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAaxF;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,MAAM,EAAE,EAAE,CAAC,EAAE,MAAM,EAAE,GAAG,MAAM,CAiBjE;AAED;;GAEG;AACH,wBAAgB,YAAY,IAAI,MAAM,CAErC;AAED;;;;GAIG;AACH,wBAAgB,OAAO,IAAI,OAAO,CAEjC;AAED;;;;;;;;;GASG;AACH,wBAAgB,iBAAiB,IAAI,OAAO,CAE3C;AAED;;GAEG;AACH,wBAAgB,QAAQ,IAAI;IAC1B,KAAK,EAAE,OAAO,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,OAAO,CAAC;IACd,QAAQ,EAAE,OAAO,CAAC;IAClB,eAAe,EAAE,MAAM,CAAC;IACxB,iBAAiB,EAAE,MAAM,CAAC;CAC3B,CAUA;AAED;;GAEG;AACH,wBAAsB,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC,CAO9C;AAED;;;;;;;GAOG;AACH,wBAAsB,oBAAoB,CAAC,UAAU,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,OAAO,CAAC,CAkBhF;AAED;;;GAGG;AACH,wBAAsB,kBAAkB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,CAG7E;AAED,mEAAmE;AACnE,wBAAgB,sBAAsB,IAAI,MAAM,CAE/C;AAED,iEAAiE;AACjE,wBAAsB,wBAAwB,IAAI,OAAO,CAAC,IAAI,CAAC,CAK9D;AAGD,qBAAa,YAAY;IACvB,OAAO,CAAC,MAAM,CAAqB;gBAEvB,MAAM,GAAE,kBAAuB;IAIrC,IAAI,IAAI,OAAO,CAAC,OAAO,CAAC;IAIxB,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC;IAKtC,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC;IAKhD,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAK/D,IAAI,SAAS,IAAI,MAAM,CAEtB;IAED,IAAI,KAAK,IAAI,OAAO,CAEnB;CACF;AAED,eAAe,YAAY,CAAC"}

package/dist/core/onnx-embedder.js

@@ -57,8 +57,13 @@ exports.similarity = similarity;
 exports.cosineSimilarity = cosineSimilarity;
 exports.getDimension = getDimension;
 exports.isReady = isReady;
+exports.isOnnxInitialized = isOnnxInitialized;
 exports.getStats = getStats;
 exports.shutdown = shutdown;
+exports.initParallelEmbedder = initParallelEmbedder;
+exports.embedBatchParallel = embedBatchParallel;
+exports.getParallelWorkerCount = getParallelWorkerCount;
+exports.shutdownParallelEmbedder = shutdownParallelEmbedder;
 const path = __importStar(require("path"));
 const fs = __importStar(require("fs"));
 const url_1 = require("url");
@@ -95,10 +100,22 @@ let loadPromise = null;
 let isInitialized = false;
 let parallelEnabled = false;
 let parallelThreshold = 4;
+// Captured at init so the bundled worker pool can reuse the loaded model bytes
+// (shared to workers via SharedArrayBuffer) instead of re-downloading per worker.
+let loadedModelBytes = null;
+let loadedTokenizerJson = null;
+let loadedMaxLength = 256;
+let bundledPool = null;
 // Default model
 const DEFAULT_MODEL = 'all-MiniLM-L6-v2';
 /**
- * Check if ONNX embedder is available (bundled files exist)
+ * Check if the ONNX embedder is *available* — i.e. the bundled WASM files are
+ * present and the embedder can be initialized.
+ *
+ * NOTE: This is a capability check, NOT a readiness check. It returns `true`
+ * before `initOnnxEmbedder()` has run (so callers can decide whether to init).
+ * To check whether the model has actually been loaded, use `isOnnxInitialized()`
+ * or `isReady()`. See https://github.com/ruvnet/RuVector/issues/523.
  */
 function isOnnxAvailable() {
     try {
@@ -143,7 +160,7 @@ async function tryInitParallel(config) {
     // Skip if explicitly disabled
     if (config.enableParallel === false)
         return false;
-    // For 'auto' or true, try to initialize
+    // 1) Optional external package (back-compat). Absent by default.
     try {
         const parallelModule = await dynamicImport('ruvector-onnx-embeddings-wasm/parallel');
         const { ParallelEmbedder } = parallelModule;
@@ -154,15 +171,44 @@ async function tryInitParallel(config) {
         parallelThreshold = config.parallelThreshold || 4;
         parallelEnabled = true;
         parallelAvailable = true;
-        console.error(`Parallel embedder ready: ${parallelEmbedder.numWorkers} workers, SIMD: ${simdAvailable}`);
+        console.error(`Parallel embedder ready (external): ${parallelEmbedder.numWorkers} workers, SIMD: ${simdAvailable}`);
         return true;
     }
-    catch (e) {
+    catch {
+        // External package not installed — fall through to the bundled pool.
+    }
+    // 2) Bundled, zero-dependency worker pool over the already-loaded model bytes.
+    // Opt-in only (enableParallel === true) so the default/'auto' path does not
+    // silently spawn worker threads for existing callers. Vectors are identical
+    // to the single-thread path (issue #523).
+    if (config.enableParallel !== true) {
         parallelAvailable = false;
-        if (config.enableParallel === true) {
-            // Only warn if explicitly requested
-            console.error(`Parallel embedder not available: ${e.message}`);
+        return false;
+    }
+    try {
+        if (!loadedModelBytes || !loadedTokenizerJson) {
+            throw new Error('model bytes unavailable for bundled pool');
         }
+        const poolUrl = (0, url_1.pathToFileURL)(path.join(__dirname, 'onnx', 'bundled-parallel.mjs')).href;
+        const { ParallelEmbedder } = await dynamicImport(poolUrl);
+        const pool = new ParallelEmbedder({
+            modelBytes: loadedModelBytes,
+            tokenizerJson: loadedTokenizerJson,
+            maxLength: loadedMaxLength,
+            dimension: embedder ? embedder.dimension() : 384,
+            numWorkers: config.numWorkers,
+        });
+        await pool.init();
+        parallelEmbedder = pool;
+        parallelThreshold = config.parallelThreshold || 4;
+        parallelEnabled = true;
+        parallelAvailable = true;
+        console.error(`Parallel embedder ready (bundled): ${pool.numWorkers} workers, SIMD: ${simdAvailable}`);
+        return true;
+    }
+    catch (e) {
+        parallelAvailable = false;
+        console.error(`Parallel embedder not available: ${e.message}`);
         return false;
     }
 }
@@ -213,6 +259,10 @@ async function initOnnxEmbedder(config = {}) {
             const modelId = config.modelId || DEFAULT_MODEL;
             console.error(`Loading ONNX model: ${modelId}...`);
             const { modelBytes, tokenizerJson, config: modelConfig } = await modelLoader.loadModel(modelId);
+            // Retain for the bundled parallel worker pool (see initParallelEmbedder).
+            loadedModelBytes = modelBytes;
+            loadedTokenizerJson = tokenizerJson;
+            loadedMaxLength = config.maxLength || modelConfig.maxLength || 256;
             // Create embedder with config
             const embedderConfig = new wasmModule.WasmEmbedderConfig()
                 .setMaxLength(config.maxLength || modelConfig.maxLength || 256)
@@ -354,11 +404,26 @@ function getDimension() {
     return embedder ? embedder.dimension() : 384;
 }
 /**
- * Check if embedder is ready
+ * Check if the embedder has been initialized (model loaded) and is ready to
+ * embed. Returns `false` until `initOnnxEmbedder()` (or the first `embed()`,
+ * which auto-initializes) has completed successfully.
  */
 function isReady() {
     return isInitialized;
 }
+/**
+ * Whether the ONNX embedder has been initialized (model loaded).
+ *
+ * Post-init counterpart to `isOnnxAvailable()` (which only checks that the
+ * bundled files exist). Named distinctly from the WASM-core `isInitialized()`
+ * export to avoid a barrel name collision. Equivalent to `isReady()`; provided
+ * as a self-documenting gate so callers can distinguish "bundled" (available)
+ * from "loaded" (initialized). See
+ * https://github.com/ruvnet/RuVector/issues/523.
+ */
+function isOnnxInitialized() {
+    return isInitialized;
+}
 /**
  * Get embedder stats including SIMD and parallel capabilities
  */
@@ -382,6 +447,56 @@ async function shutdown() {
         parallelEmbedder = null;
         parallelEnabled = false;
     }
+    await shutdownParallelEmbedder();
+}
+/**
+ * Initialize the bundled-WASM worker pool for high-throughput batch embedding
+ * (issue #523 SOTA). Self-contained — uses Node worker_threads + the bundled
+ * WASM over SharedArrayBuffer model bytes, no external dependency. Vectors are
+ * identical to the single-thread path (cosine-equivalent).
+ *
+ * @param numWorkers number of worker threads (default: min(cpus-2, 16))
+ */
+async function initParallelEmbedder(numWorkers) {
+    if (bundledPool)
+        return true;
+    if (!isInitialized)
+        await initOnnxEmbedder();
+    if (!loadedModelBytes || !loadedTokenizerJson) {
+        throw new Error('Model bytes unavailable; cannot start parallel embedder.');
+    }
+    const poolUrl = (0, url_1.pathToFileURL)(path.join(__dirname, 'onnx', 'bundled-parallel.mjs')).href;
+    const { ParallelEmbedder } = await dynamicImport(poolUrl);
+    const pool = new ParallelEmbedder({
+        modelBytes: loadedModelBytes,
+        tokenizerJson: loadedTokenizerJson,
+        maxLength: loadedMaxLength,
+        dimension: getDimension(),
+        numWorkers,
+    });
+    await pool.init();
+    bundledPool = pool;
+    return true;
+}
+/**
+ * Batch-embed via the bundled worker pool, sharded across CPU cores. Lazily
+ * starts the pool on first use. Returns embeddings in input order.
+ */
+async function embedBatchParallel(texts) {
+    if (!bundledPool)
+        await initParallelEmbedder();
+    return bundledPool.embedBatch(texts);
+}
+/** Number of active pool workers (0 if the pool isn't started). */
+function getParallelWorkerCount() {
+    return bundledPool ? bundledPool.numWorkers : 0;
+}
+/** Shut down the bundled worker pool and release its threads. */
+async function shutdownParallelEmbedder() {
+    if (bundledPool) {
+        await bundledPool.shutdown();
+        bundledPool = null;
+    }
 }
 // Export class wrapper for compatibility
 class OnnxEmbedder {

package/dist/core/onnx-optimized.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"onnx-optimized.d.ts","sourceRoot":"","sources":["../../src/core/onnx-optimized.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAcH,MAAM,WAAW,mBAAmB;IAClC,iDAAiD;IACjD,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,uDAAuD;IACvD,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,qDAAqD;IACrD,YAAY,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,GAAG,MAAM,CAAC;IACpD,sCAAsC;IACtC,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,0CAA0C;IAC1C,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,0CAA0C;IAC1C,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,iDAAiD;IACjD,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,oDAAoD;IACpD,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,qDAAqD;IACrD,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AA0HD,qBAAa,qBAAqB;IAChC,OAAO,CAAC,MAAM,CAAgC;IAC9C,OAAO,CAAC,UAAU,CAAa;IAC/B,OAAO,CAAC,QAAQ,CAAa;IAC7B,OAAO,CAAC,WAAW,CAAS;IAC5B,OAAO,CAAC,WAAW,CAA8B;IAGjD,OAAO,CAAC,cAAc,CAAiC;IACvD,OAAO,CAAC,cAAc,CAAwB;IAG9C,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,SAAS,CAAO;gBAEZ,MAAM,GAAE,mBAAwB;IAiB5C;;OAEG;IACG,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;YAWb,MAAM;IA8EpB;;OAEG;IACG,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IAiChD;;OAEG;IACG,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,YAAY,EAAE,CAAC;IAmD1D;;OAEG;IACG,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAK/D;;OAEG;IACH,gBAAgB,CAAC,CAAC,EAAE,YAAY,EAAE,CAAC,EAAE,YAAY,GAAG,MAAM;IAmB1D;;OAEG;IACH,aAAa,IAAI;QACf,SAAS,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAC;YAAC,OAAO,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,CAAA;SAAE,CAAC;QAC3E,SAAS,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAC;YAAC,OAAO,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,CAAA;SAAE,CAAC;QAC3E,SAAS,EAAE,MAAM,CAAC;QAClB,WAAW,EAAE,MAAM,CAAC;KACrB;IASD;;OAEG;IACH,UAAU,IAAI,IAAI;IAKlB;;OAEG;IACH,YAAY,IAAI,MAAM;IAItB;;OAEG;IACH,OAAO,IAAI,OAAO;IAIlB;;OAEG;IACH,SAAS,IAAI,QAAQ,CAAC,mBAAmB,CAAC;CAG3C;AAQD,wBAAgB,wBAAwB,CAAC,MAAM,CAAC,EAAE,mBAAmB,GAAG,qBAAqB,CAK5F;AAED,wBAAsB,iBAAiB,CAAC,MAAM,CAAC,EAAE,mBAAmB,GAAG,OAAO,CAAC,qBAAqB,CAAC,CAIpG;AAED,eAAe,qBAAqB,CAAC"}
+{"version":3,"file":"onnx-optimized.d.ts","sourceRoot":"","sources":["../../src/core/onnx-optimized.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAcH,MAAM,WAAW,mBAAmB;IAClC,iDAAiD;IACjD,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,uDAAuD;IACvD,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,qDAAqD;IACrD,YAAY,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,GAAG,MAAM,CAAC;IACpD,sCAAsC;IACtC,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,0CAA0C;IAC1C,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,0CAA0C;IAC1C,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,iDAAiD;IACjD,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,oDAAoD;IACpD,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,qDAAqD;IACrD,cAAc,CAAC,EAAE,MAAM,CAAC;CACzB;AA0HD,qBAAa,qBAAqB;IAChC,OAAO,CAAC,MAAM,CAAgC;IAC9C,OAAO,CAAC,UAAU,CAAa;IAC/B,OAAO,CAAC,QAAQ,CAAa;IAC7B,OAAO,CAAC,WAAW,CAAS;IAC5B,OAAO,CAAC,WAAW,CAA8B;IAGjD,OAAO,CAAC,cAAc,CAAiC;IACvD,OAAO,CAAC,cAAc,CAAwB;IAG9C,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,SAAS,CAAO;gBAEZ,MAAM,GAAE,mBAAwB;IAiB5C;;OAEG;IACG,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC;YAWb,MAAM;IA+DpB;;OAEG;IACG,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IAiChD;;OAEG;IACG,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,YAAY,EAAE,CAAC;IAmD1D;;OAEG;IACG,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC;IAK/D;;OAEG;IACH,gBAAgB,CAAC,CAAC,EAAE,YAAY,EAAE,CAAC,EAAE,YAAY,GAAG,MAAM;IAmB1D;;OAEG;IACH,aAAa,IAAI;QACf,SAAS,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAC;YAAC,OAAO,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,CAAA;SAAE,CAAC;QAC3E,SAAS,EAAE;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,MAAM,EAAE,MAAM,CAAC;YAAC,OAAO,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,CAAA;SAAE,CAAC;QAC3E,SAAS,EAAE,MAAM,CAAC;QAClB,WAAW,EAAE,MAAM,CAAC;KACrB;IASD;;OAEG;IACH,UAAU,IAAI,IAAI;IAKlB;;OAEG;IACH,YAAY,IAAI,MAAM;IAItB;;OAEG;IACH,OAAO,IAAI,OAAO;IAIlB;;OAEG;IACH,SAAS,IAAI,QAAQ,CAAC,mBAAmB,CAAC;CAG3C;AAQD,wBAAgB,wBAAwB,CAAC,MAAM,CAAC,EAAE,mBAAmB,GAAG,qBAAqB,CAK5F;AAED,wBAAsB,iBAAiB,CAAC,MAAM,CAAC,EAAE,mBAAmB,GAAG,OAAO,CAAC,qBAAqB,CAAC,CAIpG;AAED,eAAe,qBAAqB,CAAC"}

package/dist/core/onnx-optimized.js

@@ -213,34 +213,16 @@ class OptimizedOnnxEmbedder {
             }
             const loaderModule = await dynamicImport(loaderUrl);
             const { ModelLoader } = loaderModule;
-            // Select model URL based on quantization preference
+            // NOTE (issue #523): ModelLoader.loadModel() resolves the model by modelId
+            // from its own registry (currently FP32 all-MiniLM-L6-v2). The per-variant
+            // URLs in QUANTIZED_MODELS are not wired into the loader yet, so we must not
+            // log a quantization (FP16/INT8) that is not actually applied. When the
+            // loader gains variant support, thread the selected variant through to
+            // loadModel() here instead of computing an unused URL.
             const modelInfo = QUANTIZED_MODELS[this.config.modelId];
-            let modelUrl;
             if (modelInfo) {
-                if (this.config.useQuantized && this.config.quantization !== 'none') {
-                    // Try quantized version first
-                    if (this.config.quantization === 'int8' && modelInfo.int8) {
-                        modelUrl = modelInfo.int8;
-                        console.error(`Using INT8 quantized model: ${this.config.modelId}`);
-                    }
-                    else if (modelInfo.fp16) {
-                        modelUrl = modelInfo.fp16;
-                        console.error(`Using FP16 quantized model: ${this.config.modelId}`);
-                    }
-                    else {
-                        modelUrl = modelInfo.onnx;
-                        console.error(`Using FP32 model (no quantized version): ${this.config.modelId}`);
-                    }
-                }
-                else {
-                    modelUrl = modelInfo.onnx;
-                }
                 this.dimension = modelInfo.dimension;
             }
-            else {
-                // Fallback to default loader
-                modelUrl = '';
-            }
             const modelLoader = new ModelLoader({
                 cache: true,
                 cacheDir: path.join(process.env.HOME || '/tmp', '.ruvector', 'models'),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ruvector",
-  "version": "0.2.26",
+  "version": "0.2.27",
   "description": "Self-learning vector database for Node.js — hybrid search, Graph RAG, FlashAttention-3, HNSW, 50+ attention mechanisms",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -8,7 +8,7 @@
     "ruvector": "./bin/cli.js"
   },
   "scripts": {
-    "build": "tsc && mkdir -p dist/core/onnx/pkg && cp -r src/core/onnx/pkg/. dist/core/onnx/pkg/",
+    "build": "tsc && mkdir -p dist/core/onnx && cp -r src/core/onnx/. dist/core/onnx/",
     "verify-dist": "node scripts/verify-dist.js",
     "prepack": "npm run build && npm run verify-dist",
     "prepublishOnly": "npm run build && npm run verify-dist",