ruvector 0.2.30 → 0.2.32

package/dist/core/onnx-embedder.js

@@ -48,26 +48,32 @@ var __importStar = (this && this.__importStar) || (function () {
     };
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.OnnxEmbedder = void 0;
+exports.OnnxEmbedder = exports.BULK_EMBED_THRESHOLD = void 0;
 exports.isOnnxAvailable = isOnnxAvailable;
 exports.initOnnxEmbedder = initOnnxEmbedder;
 exports.embed = embed;
+exports.embedQuery = embedQuery;
+exports.embedPassage = embedPassage;
 exports.embedBatch = embedBatch;
 exports.similarity = similarity;
 exports.cosineSimilarity = cosineSimilarity;
 exports.getDimension = getDimension;
 exports.isReady = isReady;
 exports.isOnnxInitialized = isOnnxInitialized;
+exports.getActiveModelId = getActiveModelId;
+exports.getEmbedderProvenance = getEmbedderProvenance;
 exports.getStats = getStats;
 exports.shutdown = shutdown;
 exports.initParallelEmbedder = initParallelEmbedder;
 exports.embedBatchParallel = embedBatchParallel;
 exports.getParallelWorkerCount = getParallelWorkerCount;
+exports.embedBulk = embedBulk;
 exports.shutdownParallelEmbedder = shutdownParallelEmbedder;
 const path = __importStar(require("path"));
 const fs = __importStar(require("fs"));
 const url_1 = require("url");
 const module_1 = require("module");
+const embedding_provenance_1 = require("./embedding-provenance");
 // Set up ESM-compatible require for WASM module (fixes Windows/ESM compatibility)
 // The WASM bindings use module.require for Node.js crypto, this provides a fallback
 if (typeof globalThis !== 'undefined' && !globalThis.__ruvector_require) {
@@ -106,6 +112,10 @@ let loadedModelBytes = null;
 let loadedTokenizerJson = null;
 let loadedMaxLength = 256;
 let bundledPool = null;
+// ADR-210: identity of the loaded model, for prefix policies (D4) and the
+// embedding-provenance record (D0).
+let loadedModelId = null;
+let loadedNormalize = true;
 // Default model
 const DEFAULT_MODEL = 'all-MiniLM-L6-v2';
 /**
@@ -257,6 +267,8 @@ async function initOnnxEmbedder(config = {}) {
             loadedModelBytes = modelBytes;
             loadedTokenizerJson = tokenizerJson;
             loadedMaxLength = config.maxLength || modelConfig.maxLength || 256;
+            loadedModelId = modelId;
+            loadedNormalize = config.normalize !== false;
             // Create embedder with config
             const embedderConfig = new wasmModule.WasmEmbedderConfig()
                 .setMaxLength(config.maxLength || modelConfig.maxLength || 256)
@@ -301,18 +313,18 @@ async function initOnnxEmbedder(config = {}) {
     await loadPromise;
     return isInitialized;
 }
-/**
- * Generate embedding for text
- */
-async function embed(text) {
+async function embedKind(kind, text) {
     if (!isInitialized) {
         await initOnnxEmbedder();
     }
     if (!embedder) {
         throw new Error('ONNX embedder not initialized');
     }
+    // ADR-210 D4: apply the model's registered query/passage prefix. MiniLM has
+    // empty prefixes, so the default model's output is byte-identical to before.
+    const prepared = (0, embedding_provenance_1.prefixText)(loadedModelId ?? DEFAULT_MODEL, kind, text);
     const start = performance.now();
-    const embedding = embedder.embedOne(text);
+    const embedding = embedder.embedOne(prepared);
     const timeMs = performance.now() - start;
     return {
         embedding: Array.from(embedding),
@@ -320,6 +332,21 @@ async function embed(text) {
         timeMs,
     };
 }
+/**
+ * Generate embedding for text. Equivalent to `embedPassage()` (ADR-210 D4):
+ * stored/passage text is the default; use `embedQuery()` for search queries.
+ */
+async function embed(text) {
+    return embedKind('passage', text);
+}
+/** Embed a search query, applying the model's registered query prefix (D4). */
+async function embedQuery(text) {
+    return embedKind('query', text);
+}
+/** Embed a passage/document, applying the model's registered passage prefix (D4). */
+async function embedPassage(text) {
+    return embedKind('passage', text);
+}
 /**
  * Generate embeddings for multiple texts
  * Uses parallel workers automatically for batches >= parallelThreshold
@@ -331,10 +358,12 @@ async function embedBatch(texts) {
     if (!embedder) {
         throw new Error('ONNX embedder not initialized');
     }
+    // ADR-210 D4: batch embedding is the passage path (embed() === embedPassage()).
+    const prepared = texts.map(t => (0, embedding_provenance_1.prefixText)(loadedModelId ?? DEFAULT_MODEL, 'passage', t));
     const start = performance.now();
     // Use parallel workers for large batches
-    if (parallelEnabled && parallelEmbedder && texts.length >= parallelThreshold) {
-        const batchResults = await parallelEmbedder.embedBatch(texts);
+    if (parallelEnabled && parallelEmbedder && prepared.length >= parallelThreshold) {
+        const batchResults = await parallelEmbedder.embedBatch(prepared);
         const totalTime = performance.now() - start;
         const dimension = parallelEmbedder.dimension || 384;
         return batchResults.map((emb) => ({
@@ -344,11 +373,11 @@ async function embedBatch(texts) {
         }));
     }
     // Sequential fallback
-    const batchEmbeddings = embedder.embedBatch(texts);
+    const batchEmbeddings = embedder.embedBatch(prepared);
     const totalTime = performance.now() - start;
     const dimension = embedder.dimension();
     const results = [];
-    for (let i = 0; i < texts.length; i++) {
+    for (let i = 0; i < prepared.length; i++) {
         const embedding = batchEmbeddings.slice(i * dimension, (i + 1) * dimension);
         results.push({
             embedding: Array.from(embedding),
@@ -418,6 +447,26 @@ function isReady() {
 function isOnnxInitialized() {
     return isInitialized;
 }
+/** Model id of the loaded model, or null before init (ADR-210). */
+function getActiveModelId() {
+    return loadedModelId;
+}
+/**
+ * Embedding-provenance record (ADR-210 D0) describing vectors produced by the
+ * loaded ONNX embedder, or null before the model is initialized.
+ */
+function getEmbedderProvenance() {
+    if (!isInitialized)
+        return null;
+    const modelId = loadedModelId ?? DEFAULT_MODEL;
+    return {
+        embedderKind: (0, embedding_provenance_1.embedderKindForModel)(modelId),
+        modelId,
+        dimension: getDimension(),
+        normalize: loadedNormalize,
+        prefixPolicy: (0, embedding_provenance_1.getModelPrefixSpec)(modelId).prefixPolicy,
+    };
+}
 /**
  * Get embedder stats including SIMD and parallel capabilities
  */
@@ -479,12 +528,57 @@ async function initParallelEmbedder(numWorkers) {
 async function embedBatchParallel(texts) {
     if (!bundledPool)
         await initParallelEmbedder();
-    return bundledPool.embedBatch(texts);
+    // ADR-210 D4: bulk ingest is the passage path; MiniLM prefixes are empty.
+    const prepared = texts.map(t => (0, embedding_provenance_1.prefixText)(loadedModelId ?? DEFAULT_MODEL, 'passage', t));
+    return bundledPool.embedBatch(prepared);
 }
 /** Number of active pool workers (0 if the pool isn't started). */
 function getParallelWorkerCount() {
     return bundledPool ? bundledPool.numWorkers : 0;
 }
+/** Batches at or above this size route through the worker pool (ADR-210 D3). */
+exports.BULK_EMBED_THRESHOLD = 32;
+let bulkPoolFallbackWarned = false;
+/**
+ * Default bulk-embedding path (ADR-210 D3): batches of `threshold`
+ * (default 32) or more texts route through the bundled parallel worker pool
+ * — fp32 model bytes shared across workers via SharedArrayBuffer, vectors
+ * identical to the single-thread path. Smaller batches, and any batch when
+ * pool startup fails (no worker_threads, no SharedArrayBuffer), use the
+ * single-threaded batch path with one stderr note.
+ *
+ * INT8 STATUS (honest gap, ADR-210 D3): the registered int8 variants
+ * (QUANTIZED_MODELS in onnx-optimized.ts) cannot run on the bundled WASM
+ * runtime today — its graph analyzer rejects quantized MiniLM exports
+ * ("Failed analyse for node /Unsqueeze", verified against both
+ * Xenova/all-MiniLM-L6-v2 model_quantized.onnx and the official
+ * sentence-transformers model_quint8_avx2.onnx exports). Bulk ingest
+ * therefore defaults to parallel-fp32; int8 ingest needs a Rust-side
+ * runtime upgrade in the ruvector-onnx-embeddings-wasm crate (tracked as
+ * an ADR-210 follow-up). Single-query latency keeps fp32 either way.
+ */
+async function embedBulk(texts, opts = {}) {
+    if (!texts || texts.length === 0)
+        return [];
+    const threshold = opts.threshold ?? exports.BULK_EMBED_THRESHOLD;
+    if (!isInitialized) {
+        await initOnnxEmbedder();
+    }
+    if (texts.length >= threshold) {
+        try {
+            return await embedBatchParallel(texts);
+        }
+        catch (e) {
+            if (!bulkPoolFallbackWarned) {
+                bulkPoolFallbackWarned = true;
+                console.error(`ruvector: parallel bulk-embed pool unavailable (${e?.message ?? e}); ` +
+                    `using single-threaded batch embedding.`);
+            }
+        }
+    }
+    const results = await embedBatch(texts);
+    return results.map(r => r.embedding);
+}
 /** Shut down the bundled worker pool and release its threads. */
 async function shutdownParallelEmbedder() {
     if (bundledPool) {
@@ -500,10 +594,21 @@ class OnnxEmbedder {
     async init() {
         return initOnnxEmbedder(this.config);
     }
+    /** Equivalent to embedPassage() — ADR-210 D4. */
     async embed(text) {
         const result = await embed(text);
         return result.embedding;
     }
+    /** Embed a search query with the model's registered query prefix (D4). */
+    async embedQuery(text) {
+        const result = await embedQuery(text);
+        return result.embedding;
+    }
+    /** Embed a passage/document with the model's registered passage prefix (D4). */
+    async embedPassage(text) {
+        const result = await embedPassage(text);
+        return result.embedding;
+    }
     async embedBatch(texts) {
         const results = await embedBatch(texts);
         return results.map(r => r.embedding);

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

@@ -52,9 +52,16 @@ export declare class OptimizedOnnxEmbedder {
     init(): Promise<void>;
     private doInit;
     /**
-     * Embed a single text with caching
+     * Embed a single text with caching.
+     * Equivalent to `embedPassage()` — ADR-210 D4 (plain embed = passage path).
      */
     embed(text: string): Promise<Float32Array>;
+    /** Embed a search query with the model's registered query prefix (D4). */
+    embedQuery(text: string): Promise<Float32Array>;
+    /** Embed a passage/document with the model's registered passage prefix (D4). */
+    embedPassage(text: string): Promise<Float32Array>;
+    private embedKind;
+    private embedRaw;
     /**
      * Embed multiple texts with batching and caching
      */

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;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"}
+{"version":3,"file":"onnx-optimized.d.ts","sourceRoot":"","sources":["../../src/core/onnx-optimized.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAeH,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;AA0ID,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;IAkEpB;;;OAGG;IACG,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IAIhD,0EAA0E;IACpE,UAAU,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;IAIrD,gFAAgF;IAC1E,YAAY,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,CAAC;YAIzC,SAAS;YAMT,QAAQ;IAiCtB;;OAEG;IACG,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,YAAY,EAAE,CAAC;IAsD1D;;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

@@ -55,6 +55,7 @@ exports.initOptimizedOnnx = initOptimizedOnnx;
 const path = __importStar(require("path"));
 const fs = __importStar(require("fs"));
 const url_1 = require("url");
+const embedding_provenance_1 = require("./embedding-provenance");
 // Force native dynamic import
 // eslint-disable-next-line @typescript-eslint/no-implied-eval
 const dynamicImport = new Function('specifier', 'return import(specifier)');
@@ -70,6 +71,9 @@ const QUANTIZED_MODELS = {
         tokenizer: 'https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main/tokenizer.json',
         dimension: 384,
         maxLength: 256,
+        prefixPolicy: 'none',
+        queryPrefix: '',
+        passagePrefix: '',
     },
     'bge-small-en-v1.5': {
         onnx: 'https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/onnx/model.onnx',
@@ -78,6 +82,11 @@ const QUANTIZED_MODELS = {
         tokenizer: 'https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.json',
         dimension: 384,
         maxLength: 512,
+        // Query instruction recommended for short-query → long-passage retrieval;
+        // passages need no instruction (model card).
+        prefixPolicy: 'query-recommended',
+        queryPrefix: 'Represent this sentence for searching relevant passages: ',
+        passagePrefix: '',
     },
     'e5-small-v2': {
         onnx: 'https://huggingface.co/intfloat/e5-small-v2/resolve/main/onnx/model.onnx',
@@ -85,6 +94,10 @@ const QUANTIZED_MODELS = {
         tokenizer: 'https://huggingface.co/intfloat/e5-small-v2/resolve/main/tokenizer.json',
         dimension: 384,
         maxLength: 512,
+        // The model card states quality degrades without these prefixes.
+        prefixPolicy: 'required',
+        queryPrefix: 'query: ',
+        passagePrefix: 'passage: ',
     },
 };
 // ============================================================================
@@ -219,7 +232,10 @@ class OptimizedOnnxEmbedder {
             // 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];
+            // Own-property lookup only ('__proto__'-style ids must miss, ADR-210).
+            const modelInfo = Object.prototype.hasOwnProperty.call(QUANTIZED_MODELS, this.config.modelId)
+                ? QUANTIZED_MODELS[this.config.modelId]
+                : undefined;
             if (modelInfo) {
                 this.dimension = modelInfo.dimension;
             }
@@ -246,9 +262,26 @@ class OptimizedOnnxEmbedder {
         }
     }
     /**
-     * Embed a single text with caching
+     * Embed a single text with caching.
+     * Equivalent to `embedPassage()` — ADR-210 D4 (plain embed = passage path).
      */
     async embed(text) {
+        return this.embedKind('passage', text);
+    }
+    /** Embed a search query with the model's registered query prefix (D4). */
+    async embedQuery(text) {
+        return this.embedKind('query', text);
+    }
+    /** Embed a passage/document with the model's registered passage prefix (D4). */
+    async embedPassage(text) {
+        return this.embedKind('passage', text);
+    }
+    async embedKind(kind, text) {
+        // ADR-210 D4: prefix before tokenization (and before the cache key, so
+        // query and passage embeds of the same text never collide for E5/BGE).
+        return this.embedRaw((0, embedding_provenance_1.prefixText)(this.config.modelId, kind, text));
+    }
+    async embedRaw(text) {
         if (this.config.lazyInit && !this.initialized) {
             await this.init();
         }
@@ -284,17 +317,19 @@ class OptimizedOnnxEmbedder {
         if (!this.embedder) {
             throw new Error('Embedder not initialized');
         }
-        const results = new Array(texts.length);
+        // ADR-210 D4: batch embedding is the passage path (embed() === embedPassage()).
+        const prepared = texts.map(t => (0, embedding_provenance_1.prefixText)(this.config.modelId, 'passage', t));
+        const results = new Array(prepared.length);
         const uncached = [];
         // Check cache first
-        for (let i = 0; i < texts.length; i++) {
-            const cacheKey = hashString(texts[i]);
+        for (let i = 0; i < prepared.length; i++) {
+            const cacheKey = hashString(prepared[i]);
             const cached = this.embeddingCache.get(cacheKey);
             if (cached) {
                 results[i] = cached;
             }
             else {
-                uncached.push({ index: i, text: texts[i] });
+                uncached.push({ index: i, text: prepared[i] });
             }
         }
         // If all cached, return immediately

package/package.json

@@ -1,18 +1,18 @@
 {
   "name": "ruvector",
-  "version": "0.2.30",
-  "description": "Self-learning vector database for Node.js — hybrid search, Graph RAG, FlashAttention-3, HNSW, 50+ attention mechanisms",
+  "version": "0.2.32",
+  "description": "Self-learning vector database for Node.js \u2014 hybrid search, Graph RAG, FlashAttention-3, HNSW, 50+ attention mechanisms",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
   "bin": {
     "ruvector": "./bin/cli.js"
   },
   "scripts": {
-    "build": "tsc && mkdir -p dist/core/onnx && cp -r src/core/onnx/. dist/core/onnx/",
+    "build": "tsc && node -e \"require('fs').cpSync('src/core/onnx','dist/core/onnx',{recursive:true})\"",
     "verify-dist": "node scripts/verify-dist.js",
     "prepack": "npm run build && npm run verify-dist",
     "prepublishOnly": "npm run build && npm run verify-dist",
-    "test": "node test/integration.js && node test/cli-commands.js && node test/sigterm-cleanup.js"
+    "test": "node test/integration.js && node test/cli-commands.js && node test/db-workflow.js && node test/sigterm-cleanup.js && node test/mcp-policy.js && node test/startup-budget.js"
   },
   "keywords": [
     "vector",
@@ -82,7 +82,8 @@
     "ora": "^5.4.1"
   },
   "optionalDependencies": {
-    "@ruvector/rvf": "^0.1.0"
+    "@ruvector/rvf": "^0.1.0",
+    "@ruvector/tiny-dancer": "^0.1.22"
   },
   "devDependencies": {
     "@types/node": "^20.10.5",