agentic-flow 2.0.1-alpha.14 → 2.0.1-alpha.15

package/dist/intelligence/EmbeddingService.js

@@ -1,14 +1,37 @@
 /**
  * EmbeddingService - Unified embedding interface for agentic-flow
  *
- * Supports two backends:
- * - Simple: Hash-based, ~0.01ms, not semantic
- * - ONNX: Real neural embeddings via ruvector-onnx-embeddings-wasm, ~10-50ms, semantic
+ * Uses ruvector@0.1.61+ for ONNX embeddings with:
+ * - SIMD128 acceleration (6x faster)
+ * - Parallel worker threads (7 workers)
+ * - all-MiniLM-L6-v2 model (384 dimensions)
  *
  * Configure via:
- * - AGENTIC_FLOW_EMBEDDINGS=simple|onnx (default: simple)
+ * - AGENTIC_FLOW_EMBEDDINGS=simple|onnx|auto (default: auto)
  * - AGENTIC_FLOW_EMBEDDING_MODEL=all-MiniLM-L6-v2 (default)
  */
+// ONNX availability cache
+let onnxAvailable = null;
+let ruvectorModule = null;
+/**
+ * Detect ONNX/SIMD support by loading ruvector
+ */
+async function detectOnnx() {
+    if (onnxAvailable !== null) {
+        return onnxAvailable;
+    }
+    try {
+        const mod = await import('ruvector');
+        ruvectorModule = mod;
+        onnxAvailable = mod.isOnnxAvailable?.() ?? false;
+        return onnxAvailable;
+    }
+    catch (error) {
+        // Ruvector loading failed - fall back to simple embeddings
+        onnxAvailable = false;
+        return false;
+    }
+}
 // Simple LRU cache for embeddings
 class EmbeddingCache {
     cache = new Map();
@@ -45,13 +68,12 @@ class EmbeddingCache {
 export class EmbeddingService {
     static instance = null;
     backend;
+    effectiveBackend = null;
     dimension;
     modelName;
     // ONNX state
-    embedder = null;
-    wasmModule = null;
-    loadingPromise = null;
     modelLoaded = false;
+    loadingPromise = null;
     // Stats
     totalEmbeddings = 0;
     totalLatencyMs = 0;
@@ -59,10 +81,13 @@ export class EmbeddingService {
     // Cache
     cache;
     cacheEnabled;
+    // Corpus for search operations
+    corpus = { texts: [], embeddings: [] };
     constructor() {
-        this.backend = process.env.AGENTIC_FLOW_EMBEDDINGS || 'simple';
+        // Default to 'auto' which will detect ONNX and use it if available
+        this.backend = process.env.AGENTIC_FLOW_EMBEDDINGS || 'auto';
         this.modelName = process.env.AGENTIC_FLOW_EMBEDDING_MODEL || 'all-MiniLM-L6-v2';
-        this.dimension = this.backend === 'onnx' ? 384 : 256; // ONNX uses 384, simple uses 256
+        this.dimension = 256; // Will be updated when ONNX loads (384)
         this.cacheEnabled = process.env.AGENTIC_FLOW_EMBEDDING_CACHE !== 'false';
         this.cache = new EmbeddingCache(1000);
     }
@@ -73,11 +98,40 @@ export class EmbeddingService {
         return EmbeddingService.instance;
     }
     /**
-     * Get current backend
+     * Resolve the effective backend based on ONNX detection
+     */
+    async resolveBackend() {
+        if (this.effectiveBackend) {
+            return this.effectiveBackend;
+        }
+        if (this.backend === 'auto') {
+            const hasOnnx = await detectOnnx();
+            this.effectiveBackend = hasOnnx ? 'onnx' : 'simple';
+            if (hasOnnx) {
+                this.dimension = 384; // all-MiniLM-L6-v2 dimension
+            }
+        }
+        else {
+            this.effectiveBackend = this.backend;
+            if (this.backend === 'onnx') {
+                await detectOnnx(); // Ensure module is loaded
+                this.dimension = 384;
+            }
+        }
+        return this.effectiveBackend;
+    }
+    /**
+     * Get configured backend (may be 'auto')
      */
     getBackend() {
         return this.backend;
     }
+    /**
+     * Get effective backend after detection
+     */
+    getEffectiveBackend() {
+        return this.effectiveBackend || this.backend;
+    }
     /**
      * Get embedding dimension
      */
@@ -90,54 +144,9 @@ export class EmbeddingService {
     isModelLoaded() {
         return this.modelLoaded;
     }
-    /**
-     * Initialize ONNX embedder (lazy load on first use)
-     */
-    async initOnnx() {
-        if (this.modelLoaded)
-            return;
-        if (this.loadingPromise)
-            return this.loadingPromise;
-        this.loadingPromise = this._loadOnnxModel();
-        await this.loadingPromise;
-    }
-    async _loadOnnxModel() {
-        try {
-            console.log(`[EmbeddingService] Loading ONNX model: ${this.modelName}`);
-            const startTime = performance.now();
-            // Import WASM module (auto-initializes with --experimental-wasm-modules)
-            const wasmModule = await import('ruvector-onnx-embeddings-wasm');
-            this.wasmModule = wasmModule;
-            // Import loader for model fetching
-            const loaderModule = await import('ruvector-onnx-embeddings-wasm/loader.js');
-            const { ModelLoader, MODELS } = loaderModule;
-            const modelConfig = MODELS[this.modelName];
-            if (!modelConfig) {
-                throw new Error(`Unknown model: ${this.modelName}`);
-            }
-            // Load model files
-            const loader = new ModelLoader();
-            const { modelBytes, tokenizerJson } = await loader.loadModel(this.modelName);
-            // Create embedder with configuration
-            const embedderConfig = new wasmModule.WasmEmbedderConfig()
-                .setMaxLength(modelConfig.maxLength)
-                .setNormalize(true)
-                .setPooling(0); // Mean pooling
-            this.embedder = wasmModule.WasmEmbedder.withConfig(modelBytes, tokenizerJson, embedderConfig);
-            this.dimension = this.embedder.dimension();
-            this.modelLoaded = true;
-            const loadTime = performance.now() - startTime;
-            console.log(`[EmbeddingService] Model loaded in ${loadTime.toFixed(0)}ms, dim=${this.dimension}`);
-        }
-        catch (error) {
-            console.error('[EmbeddingService] Failed to load ONNX model:', error);
-            console.log('[EmbeddingService] Falling back to simple embeddings');
-            this.backend = 'simple';
-            this.dimension = 256;
-        }
-    }
     /**
      * Generate embedding for text
+     * Auto-detects ONNX and uses it if available (default behavior)
      */
     async embed(text) {
         const startTime = performance.now();
@@ -149,11 +158,14 @@ export class EmbeddingService {
                 return cached;
             }
         }
+        // Resolve backend (handles 'auto' mode)
+        const effectiveBackend = await this.resolveBackend();
         let embedding;
-        if (this.backend === 'onnx') {
-            await this.initOnnx();
-            if (this.embedder) {
-                embedding = this.embedder.embedOne(text);
+        if (effectiveBackend === 'onnx' && ruvectorModule) {
+            const result = await ruvectorModule.embed(text);
+            if (result?.embedding) {
+                embedding = result.embedding;
+                this.modelLoaded = true;
             }
             else {
                 embedding = this.simpleEmbed(text);
@@ -172,38 +184,223 @@ export class EmbeddingService {
         return embedding;
     }
     /**
-     * Generate embeddings for multiple texts
+     * Generate embeddings for multiple texts (batch processing with parallel workers)
+     * Batch processing provides significant speedup with parallel ONNX workers
      */
     async embedBatch(texts) {
-        if (this.backend === 'onnx' && this.embedder) {
-            await this.initOnnx();
-            if (this.embedder) {
-                const batchEmbedding = this.embedder.embedBatch(texts);
-                const dim = this.dimension;
-                const embeddings = [];
-                for (let i = 0; i < texts.length; i++) {
-                    embeddings.push(batchEmbedding.slice(i * dim, (i + 1) * dim));
+        const startTime = performance.now();
+        // Check cache for all texts first
+        if (this.cacheEnabled) {
+            const cachedResults = texts.map(t => this.cache.get(t) || null);
+            const allCached = cachedResults.every(r => r !== null);
+            if (allCached) {
+                this.cacheHits += texts.length;
+                return cachedResults;
+            }
+        }
+        // Resolve backend
+        const effectiveBackend = await this.resolveBackend();
+        if (effectiveBackend === 'onnx' && ruvectorModule) {
+            const result = await ruvectorModule.embedBatch(texts);
+            if (result?.embeddings && result.embeddings.length === texts.length) {
+                const embeddings = result.embeddings;
+                // Cache individual embeddings
+                if (this.cacheEnabled) {
+                    for (let i = 0; i < texts.length; i++) {
+                        this.cache.set(texts[i], embeddings[i]);
+                    }
                 }
+                // Update stats
+                this.totalEmbeddings += texts.length;
+                this.totalLatencyMs += performance.now() - startTime;
+                this.modelLoaded = true;
                 return embeddings;
             }
         }
-        // Fall back to sequential
+        // Fall back to sequential for simple backend
         return Promise.all(texts.map(t => this.embed(t)));
     }
     /**
      * Compute similarity between two texts
      */
     async similarity(text1, text2) {
-        if (this.backend === 'onnx') {
-            await this.initOnnx();
-            if (this.embedder) {
-                return this.embedder.similarity(text1, text2);
-            }
+        const effectiveBackend = await this.resolveBackend();
+        if (effectiveBackend === 'onnx' && ruvectorModule) {
+            const result = await ruvectorModule.similarity(text1, text2);
+            return result.similarity;
         }
         // Fall back to embedding + cosine
         const [e1, e2] = await Promise.all([this.embed(text1), this.embed(text2)]);
         return this.cosineSimilarity(e1, e2);
     }
+    /**
+     * Compute NxN similarity matrix for a list of texts
+     * Uses parallel workers for ONNX backend
+     */
+    async similarityMatrix(texts) {
+        const embeddings = await this.embedBatch(texts);
+        const n = texts.length;
+        const matrix = Array(n).fill(null).map(() => Array(n).fill(0));
+        for (let i = 0; i < n; i++) {
+            matrix[i][i] = 1.0; // Self-similarity
+            for (let j = i + 1; j < n; j++) {
+                const sim = this.cosineSimilarity(embeddings[i], embeddings[j]);
+                matrix[i][j] = sim;
+                matrix[j][i] = sim; // Symmetric
+            }
+        }
+        return matrix;
+    }
+    /**
+     * Build a corpus for semantic search
+     */
+    async buildCorpus(texts) {
+        this.corpus.texts = texts;
+        this.corpus.embeddings = await this.embedBatch(texts);
+    }
+    /**
+     * Semantic search against the corpus
+     * Returns top-k most similar texts
+     */
+    async semanticSearch(query, topK = 5) {
+        if (this.corpus.texts.length === 0) {
+            throw new Error('Corpus not built. Call buildCorpus() first.');
+        }
+        const queryEmbedding = await this.embed(query);
+        const results = [];
+        for (let i = 0; i < this.corpus.texts.length; i++) {
+            const sim = this.cosineSimilarity(queryEmbedding, this.corpus.embeddings[i]);
+            results.push({
+                text: this.corpus.texts[i],
+                index: i,
+                similarity: sim,
+            });
+        }
+        // Sort by similarity (descending) and return top-k
+        results.sort((a, b) => b.similarity - a.similarity);
+        return results.slice(0, topK);
+    }
+    /**
+     * Find near-duplicate texts in a list
+     * Groups texts with similarity above threshold
+     */
+    async findDuplicates(texts, threshold = 0.9) {
+        const embeddings = await this.embedBatch(texts);
+        const n = texts.length;
+        const visited = new Set();
+        const groups = [];
+        for (let i = 0; i < n; i++) {
+            if (visited.has(i))
+                continue;
+            const group = {
+                indices: [i],
+                texts: [texts[i]],
+                similarity: 1.0,
+            };
+            for (let j = i + 1; j < n; j++) {
+                if (visited.has(j))
+                    continue;
+                const sim = this.cosineSimilarity(embeddings[i], embeddings[j]);
+                if (sim >= threshold) {
+                    group.indices.push(j);
+                    group.texts.push(texts[j]);
+                    group.similarity = Math.min(group.similarity, sim);
+                    visited.add(j);
+                }
+            }
+            if (group.indices.length > 1) {
+                visited.add(i);
+                groups.push(group);
+            }
+        }
+        return groups;
+    }
+    /**
+     * K-means clustering of texts
+     * Returns cluster assignments and centroids
+     */
+    async clusterTexts(texts, k = 3, maxIterations = 100) {
+        const embeddings = await this.embedBatch(texts);
+        const n = texts.length;
+        const dim = this.dimension;
+        // Initialize centroids randomly (copy to new ArrayBuffer for consistent typing)
+        const centroidIndices = new Set();
+        while (centroidIndices.size < k && centroidIndices.size < n) {
+            centroidIndices.add(Math.floor(Math.random() * n));
+        }
+        let centroids = Array.from(centroidIndices).map(i => {
+            const copy = new Float32Array(dim);
+            copy.set(embeddings[i]);
+            return copy;
+        });
+        let clusters = new Array(n).fill(0);
+        for (let iter = 0; iter < maxIterations; iter++) {
+            // Assign points to nearest centroid
+            const newClusters = embeddings.map(emb => {
+                let bestCluster = 0;
+                let bestSim = -Infinity;
+                for (let c = 0; c < k; c++) {
+                    const sim = this.cosineSimilarity(emb, centroids[c]);
+                    if (sim > bestSim) {
+                        bestSim = sim;
+                        bestCluster = c;
+                    }
+                }
+                return bestCluster;
+            });
+            // Check convergence
+            const changed = newClusters.some((c, i) => c !== clusters[i]);
+            clusters = newClusters;
+            if (!changed)
+                break;
+            // Update centroids
+            const newCentroids = [];
+            for (let c = 0; c < k; c++) {
+                newCentroids.push(new Float32Array(dim));
+            }
+            const counts = new Array(k).fill(0);
+            for (let i = 0; i < n; i++) {
+                const c = clusters[i];
+                counts[c]++;
+                for (let d = 0; d < dim; d++) {
+                    newCentroids[c][d] += embeddings[i][d];
+                }
+            }
+            // Normalize centroids
+            for (let c = 0; c < k; c++) {
+                if (counts[c] > 0) {
+                    let norm = 0;
+                    for (let d = 0; d < dim; d++) {
+                        newCentroids[c][d] /= counts[c];
+                        norm += newCentroids[c][d] * newCentroids[c][d];
+                    }
+                    norm = Math.sqrt(norm) || 1;
+                    for (let d = 0; d < dim; d++) {
+                        newCentroids[c][d] /= norm;
+                    }
+                }
+            }
+            centroids = newCentroids;
+        }
+        return { clusters, centroids };
+    }
+    /**
+     * Stream embeddings for large batches (memory efficient)
+     * Yields embeddings one at a time
+     */
+    async *streamEmbed(texts, batchSize = 32) {
+        for (let i = 0; i < texts.length; i += batchSize) {
+            const batch = texts.slice(i, i + batchSize);
+            const embeddings = await this.embedBatch(batch);
+            for (let j = 0; j < batch.length; j++) {
+                yield {
+                    index: i + j,
+                    text: batch[j],
+                    embedding: embeddings[j],
+                };
+            }
+        }
+    }
     /**
      * Simple hash-based embedding (fast, not semantic)
      */
@@ -231,8 +428,8 @@ export class EmbeddingService {
      * Compute cosine similarity between two embeddings
      */
     cosineSimilarity(a, b) {
-        if (this.wasmModule) {
-            return this.wasmModule.cosineSimilarity(a, b);
+        if (ruvectorModule?.cosineSimilarity) {
+            return ruvectorModule.cosineSimilarity(a, b);
         }
         // JS fallback
         let dot = 0;
@@ -249,16 +446,20 @@ export class EmbeddingService {
      * Get statistics
      */
     getStats() {
+        const effective = this.effectiveBackend || this.backend;
+        const ruvectorStats = ruvectorModule?.getStats?.() || {};
         return {
             backend: this.backend,
+            effectiveBackend: effective,
             dimension: this.dimension,
             totalEmbeddings: this.totalEmbeddings,
             totalLatencyMs: this.totalLatencyMs,
             avgLatencyMs: this.totalEmbeddings > 0 ? this.totalLatencyMs / this.totalEmbeddings : 0,
             cacheHits: this.cacheHits,
             modelLoaded: this.modelLoaded,
-            modelName: this.backend === 'onnx' ? this.modelName : undefined,
-            simdAvailable: this.wasmModule?.simd_available?.() ?? undefined,
+            modelName: effective === 'onnx' ? this.modelName : undefined,
+            simdAvailable: ruvectorStats.simdAvailable ?? onnxAvailable,
+            parallelWorkers: ruvectorStats.workerCount ?? undefined,
         };
     }
     /**
@@ -267,14 +468,30 @@ export class EmbeddingService {
     clearCache() {
         this.cache.clear();
     }
+    /**
+     * Clear corpus
+     */
+    clearCorpus() {
+        this.corpus = { texts: [], embeddings: [] };
+    }
+    /**
+     * Shutdown (cleanup workers)
+     */
+    async shutdown() {
+        if (ruvectorModule?.shutdown) {
+            await ruvectorModule.shutdown();
+        }
+    }
     /**
      * Reset instance (for testing)
      */
-    static reset() {
-        if (EmbeddingService.instance?.embedder) {
-            EmbeddingService.instance.embedder.free();
+    static async reset() {
+        if (EmbeddingService.instance) {
+            await EmbeddingService.instance.shutdown();
         }
         EmbeddingService.instance = null;
+        onnxAvailable = null;
+        ruvectorModule = null;
     }
 }
 // Export singleton getter
@@ -294,4 +511,16 @@ export async function textSimilarity(text1, text2) {
 export function simpleEmbed(text, dim = 256) {
     return getEmbeddingService().simpleEmbed(text, dim);
 }
+export async function similarityMatrix(texts) {
+    return getEmbeddingService().similarityMatrix(texts);
+}
+export async function semanticSearch(query, topK = 5) {
+    return getEmbeddingService().semanticSearch(query, topK);
+}
+export async function findDuplicates(texts, threshold = 0.9) {
+    return getEmbeddingService().findDuplicates(texts, threshold);
+}
+export async function clusterTexts(texts, k = 3) {
+    return getEmbeddingService().clusterTexts(texts, k);
+}
 //# sourceMappingURL=EmbeddingService.js.map

package/dist/intelligence/EmbeddingService.js.map

@@ -1 +1 @@
-{"version":3,"file":"EmbeddingService.js","sourceRoot":"","sources":["../../src/intelligence/EmbeddingService.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AA+DH,kCAAkC;AAClC,MAAM,cAAc;IACV,KAAK,GAA8B,IAAI,GAAG,EAAE,CAAC;IAC7C,OAAO,CAAS;IAExB,YAAY,UAAkB,IAAI;QAChC,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;IACzB,CAAC;IAED,GAAG,CAAC,GAAW;QACb,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAClC,IAAI,KAAK,EAAE,CAAC;YACV,mCAAmC;YACnC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;YACvB,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;QAC7B,CAAC;QACD,OAAO,KAAK,CAAC;IACf,CAAC;IAED,GAAG,CAAC,GAAW,EAAE,KAAmB;QAClC,IAAI,IAAI,CAAC,KAAK,CAAC,IAAI,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACpC,8BAA8B;YAC9B,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC;YAChD,IAAI,QAAQ,EAAE,CAAC;gBACb,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;YAC9B,CAAC;QACH,CAAC;QACD,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;IAC7B,CAAC;IAED,KAAK;QACH,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,CAAC;IACrB,CAAC;IAED,IAAI,IAAI;QACN,OAAO,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC;IACzB,CAAC;CACF;AAED,MAAM,OAAO,gBAAgB;IACnB,MAAM,CAAC,QAAQ,GAA4B,IAAI,CAAC;IAEhD,OAAO,CAAmB;IAC1B,SAAS,CAAS;IAClB,SAAS,CAAS;IAE1B,aAAa;IACL,QAAQ,GAAwB,IAAI,CAAC;IACrC,UAAU,GAAsB,IAAI,CAAC;IACrC,cAAc,GAAyB,IAAI,CAAC;IAC5C,WAAW,GAAY,KAAK,CAAC;IAErC,QAAQ;IACA,eAAe,GAAW,CAAC,CAAC;IAC5B,cAAc,GAAW,CAAC,CAAC;IAC3B,SAAS,GAAW,CAAC,CAAC;IAE9B,QAAQ;IACA,KAAK,CAAiB;IACtB,YAAY,CAAU;IAE9B;QACE,IAAI,CAAC,OAAO,GAAI,OAAO,CAAC,GAAG,CAAC,uBAA4C,IAAI,QAAQ,CAAC;QACrF,IAAI,CAAC,SAAS,GAAG,OAAO,CAAC,GAAG,CAAC,4BAA4B,IAAI,kBAAkB,CAAC;QAChF,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,OAAO,KAAK,MAAM,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,iCAAiC;QACvF,IAAI,CAAC,YAAY,GAAG,OAAO,CAAC,GAAG,CAAC,4BAA4B,KAAK,OAAO,CAAC;QACzE,IAAI,CAAC,KAAK,GAAG,IAAI,cAAc,CAAC,IAAI,CAAC,CAAC;IACxC,CAAC;IAED,MAAM,CAAC,WAAW;QAChB,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,CAAC;YAC/B,gBAAgB,CAAC,QAAQ,GAAG,IAAI,gBAAgB,EAAE,CAAC;QACrD,CAAC;QACD,OAAO,gBAAgB,CAAC,QAAQ,CAAC;IACnC,CAAC;IAED;;OAEG;IACH,UAAU;QACR,OAAO,IAAI,CAAC,OAAO,CAAC;IACtB,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO,IAAI,CAAC,SAAS,CAAC;IACxB,CAAC;IAED;;OAEG;IACH,aAAa;QACX,OAAO,IAAI,CAAC,WAAW,CAAC;IAC1B,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,QAAQ;QACZ,IAAI,IAAI,CAAC,WAAW;YAAE,OAAO;QAC7B,IAAI,IAAI,CAAC,cAAc;YAAE,OAAO,IAAI,CAAC,cAAc,CAAC;QAEpD,IAAI,CAAC,cAAc,GAAG,IAAI,CAAC,cAAc,EAAE,CAAC;QAC5C,MAAM,IAAI,CAAC,cAAc,CAAC;IAC5B,CAAC;IAEO,KAAK,CAAC,cAAc;QAC1B,IAAI,CAAC;YACH,OAAO,CAAC,GAAG,CAAC,0CAA0C,IAAI,CAAC,SAAS,EAAE,CAAC,CAAC;YACxE,MAAM,SAAS,GAAG,WAAW,CAAC,GAAG,EAAE,CAAC;YAEpC,yEAAyE;YACzE,MAAM,UAAU,GAAG,MAAM,MAAM,CAAC,+BAA+B,CAAC,CAAC;YACjE,IAAI,CAAC,UAAU,GAAG,UAAmC,CAAC;YAEtD,mCAAmC;YACnC,MAAM,YAAY,GAAG,MAAM,MAAM,CAAC,yCAAyC,CAAC,CAAC;YAC7E,MAAM,EAAE,WAAW,EAAE,MAAM,EAAE,GAAG,YAAY,CAAC;YAE7C,MAAM,WAAW,GAAG,MAAM,CAAC,IAAI,CAAC,SAAS,CAAgB,CAAC;YAC1D,IAAI,CAAC,WAAW,EAAE,CAAC;gBACjB,MAAM,IAAI,KAAK,CAAC,kBAAkB,IAAI,CAAC,SAAS,EAAE,CAAC,CAAC;YACtD,CAAC;YAED,mBAAmB;YACnB,MAAM,MAAM,GAAG,IAAI,WAAW,EAAE,CAAC;YACjC,MAAM,EAAE,UAAU,EAAE,aAAa,EAAE,GAAG,MAAM,MAAM,CAAC,SAAS,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YAE7E,qCAAqC;YACrC,MAAM,cAAc,GAAG,IAAK,UAAkB,CAAC,kBAAkB,EAAE;iBAChE,YAAY,CAAC,WAAW,CAAC,SAAS,CAAC;iBACnC,YAAY,CAAC,IAAI,CAAC;iBAClB,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,eAAe;YAEjC,IAAI,CAAC,QAAQ,GAAI,UAAkB,CAAC,YAAY,CAAC,UAAU,CACzD,UAAU,EACV,aAAa,EACb,cAAc,CACf,CAAC;YAEF,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,QAAS,CAAC,SAAS,EAAE,CAAC;YAC5C,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC;YAExB,MAAM,QAAQ,GAAG,WAAW,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;YAC/C,OAAO,CAAC,GAAG,CAAC,sCAAsC,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,WAAW,IAAI,CAAC,SAAS,EAAE,CAAC,CAAC;QACpG,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,CAAC,KAAK,CAAC,+CAA+C,EAAE,KAAK,CAAC,CAAC;YACtE,OAAO,CAAC,GAAG,CAAC,sDAAsD,CAAC,CAAC;YACpE,IAAI,CAAC,OAAO,GAAG,QAAQ,CAAC;YACxB,IAAI,CAAC,SAAS,GAAG,GAAG,CAAC;QACvB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,KAAK,CAAC,IAAY;QACtB,MAAM,SAAS,GAAG,WAAW,CAAC,GAAG,EAAE,CAAC;QAEpC,cAAc;QACd,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;YACtB,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;YACpC,IAAI,MAAM,EAAE,CAAC;gBACX,IAAI,CAAC,SAAS,EAAE,CAAC;gBACjB,OAAO,MAAM,CAAC;YAChB,CAAC;QACH,CAAC;QAED,IAAI,SAAuB,CAAC;QAE5B,IAAI,IAAI,CAAC,OAAO,KAAK,MAAM,EAAE,CAAC;YAC5B,MAAM,IAAI,CAAC,QAAQ,EAAE,CAAC;YACtB,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAClB,SAAS,GAAG,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;YAC3C,CAAC;iBAAM,CAAC;gBACN,SAAS,GAAG,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;YACrC,CAAC;QACH,CAAC;aAAM,CAAC;YACN,SAAS,GAAG,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;QACrC,CAAC;QAED,eAAe;QACf,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,IAAI,CAAC,cAAc,IAAI,WAAW,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;QAErD,eAAe;QACf,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;YACtB,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;QAClC,CAAC;QAED,OAAO,SAAS,CAAC;IACnB,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,UAAU,CAAC,KAAe;QAC9B,IAAI,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;YAC7C,MAAM,IAAI,CAAC,QAAQ,EAAE,CAAC;YACtB,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAClB,MAAM,cAAc,GAAG,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;gBACvD,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS,CAAC;gBAC3B,MAAM,UAAU,GAAmB,EAAE,CAAC;gBACtC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;oBACtC,UAAU,CAAC,IAAI,CAAC,cAAc,CAAC,KAAK,CAAC,CAAC,GAAG,GAAG,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC;gBAChE,CAAC;gBACD,OAAO,UAAU,CAAC;YACpB,CAAC;QACH,CAAC;QAED,0BAA0B;QAC1B,OAAO,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,UAAU,CAAC,KAAa,EAAE,KAAa;QAC3C,IAAI,IAAI,CAAC,OAAO,KAAK,MAAM,EAAE,CAAC;YAC5B,MAAM,IAAI,CAAC,QAAQ,EAAE,CAAC;YACtB,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;gBAClB,OAAO,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;YAChD,CAAC;QACH,CAAC;QAED,kCAAkC;QAClC,MAAM,CAAC,EAAE,EAAE,EAAE,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAC3E,OAAO,IAAI,CAAC,gBAAgB,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC;IACvC,CAAC;IAED;;OAEG;IACH,WAAW,CAAC,IAAY,EAAE,MAAc,GAAG;QACzC,MAAM,SAAS,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC;QAExC,0CAA0C;QAC1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACrC,MAAM,IAAI,GAAG,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC;YAChC,SAAS,CAAC,CAAC,GAAG,GAAG,CAAC,IAAI,IAAI,GAAG,GAAG,CAAC;YACjC,SAAS,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC,IAAI,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC;YAC/C,SAAS,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC,IAAI,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC;QAClD,CAAC;QAED,YAAY;QACZ,IAAI,IAAI,GAAG,CAAC,CAAC;QACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7B,IAAI,IAAI,SAAS,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC;QACtC,CAAC;QACD,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAC5B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7B,SAAS,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;QACvB,CAAC;QAED,OAAO,SAAS,CAAC;IACnB,CAAC;IAED;;OAEG;IACH,gBAAgB,CAAC,CAAe,EAAE,CAAe;QAC/C,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;YACpB,OAAO,IAAI,CAAC,UAAU,CAAC,gBAAgB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QAChD,CAAC;QAED,cAAc;QACd,IAAI,GAAG,GAAG,CAAC,CAAC;QACZ,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAClC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACnB,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;QACvB,CAAC;QACD,OAAO,GAAG,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC;IAC1D,CAAC;IAED;;OAEG;IACH,QAAQ;QACN,OAAO;YACL,OAAO,EAAE,IAAI,CAAC,OAAO;YACrB,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,eAAe,EAAE,IAAI,CAAC,eAAe;YACrC,cAAc,EAAE,IAAI,CAAC,cAAc;YACnC,YAAY,EAAE,IAAI,CAAC,eAAe,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,GAAG,IAAI,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;YACvF,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,WAAW,EAAE,IAAI,CAAC,WAAW;YAC7B,SAAS,EAAE,IAAI,CAAC,OAAO,KAAK,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS;YAC/D,aAAa,EAAE,IAAI,CAAC,UAAU,EAAE,cAAc,EAAE,EAAE,IAAI,SAAS;SAChE,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,UAAU;QACR,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,CAAC;IACrB,CAAC;IAED;;OAEG;IACH,MAAM,CAAC,KAAK;QACV,IAAI,gBAAgB,CAAC,QAAQ,EAAE,QAAQ,EAAE,CAAC;YACxC,gBAAgB,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC;QAC5C,CAAC;QACD,gBAAgB,CAAC,QAAQ,GAAG,IAAI,CAAC;IACnC,CAAC;;AAGH,0BAA0B;AAC1B,MAAM,UAAU,mBAAmB;IACjC,OAAO,gBAAgB,CAAC,WAAW,EAAE,CAAC;AACxC,CAAC;AAED,+BAA+B;AAC/B,MAAM,CAAC,KAAK,UAAU,KAAK,CAAC,IAAY;IACtC,OAAO,mBAAmB,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;AAC3C,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,UAAU,CAAC,KAAe;IAC9C,OAAO,mBAAmB,EAAE,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;AACjD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,KAAa,EAAE,KAAa;IAC/D,OAAO,mBAAmB,EAAE,CAAC,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;AACxD,CAAC;AAED,MAAM,UAAU,WAAW,CAAC,IAAY,EAAE,MAAc,GAAG;IACzD,OAAO,mBAAmB,EAAE,CAAC,WAAW,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;AACtD,CAAC","sourcesContent":["/**\n * EmbeddingService - Unified embedding interface for agentic-flow\n *\n * Supports two backends:\n * - Simple: Hash-based, ~0.01ms, not semantic\n * - ONNX: Real neural embeddings via ruvector-onnx-embeddings-wasm, ~10-50ms, semantic\n *\n * Configure via:\n * - AGENTIC_FLOW_EMBEDDINGS=simple|onnx (default: simple)\n * - AGENTIC_FLOW_EMBEDDING_MODEL=all-MiniLM-L6-v2 (default)\n */\n\n// Types for ONNX Runtime\ninterface OnnxSession {\n  run(feeds: Record<string, any>): Promise<Record<string, any>>;\n  release(): void;\n}\n\ninterface Tokenizer {\n  encode(text: string, addSpecialTokens?: boolean): { ids: number[]; attentionMask: number[] };\n}\n\n// Types for the WASM module\ninterface WasmEmbedder {\n  embedOne(text: string): Float32Array;\n  embedBatch(texts: string[]): Float32Array;\n  similarity(text1: string, text2: string): number;\n  dimension(): number;\n  maxLength(): number;\n  free(): void;\n}\n\ninterface WasmModule {\n  default: () => Promise<void>;\n  WasmEmbedder: {\n    withConfig(modelBytes: Uint8Array, tokenizerJson: string, config: WasmEmbedderConfig): WasmEmbedder;\n  };\n  WasmEmbedderConfig: new () => WasmEmbedderConfig;\n  cosineSimilarity(a: Float32Array, b: Float32Array): number;\n  simd_available(): boolean;\n  version(): string;\n}\n\ninterface WasmEmbedderConfig {\n  setMaxLength(maxLength: number): WasmEmbedderConfig;\n  setNormalize(normalize: boolean): WasmEmbedderConfig;\n  setPooling(pooling: number): WasmEmbedderConfig;\n}\n\ninterface ModelConfig {\n  name: string;\n  dimension: number;\n  maxLength: number;\n  size: string;\n  description: string;\n  model: string;\n  tokenizer: string;\n}\n\nexport type EmbeddingBackend = 'simple' | 'onnx' | 'onnx-native';\n\nexport interface EmbeddingStats {\n  backend: EmbeddingBackend;\n  dimension: number;\n  totalEmbeddings: number;\n  totalLatencyMs: number;\n  avgLatencyMs: number;\n  cacheHits: number;\n  modelLoaded: boolean;\n  modelName?: string;\n  simdAvailable?: boolean;\n}\n\n// Simple LRU cache for embeddings\nclass EmbeddingCache {\n  private cache: Map<string, Float32Array> = new Map();\n  private maxSize: number;\n\n  constructor(maxSize: number = 1000) {\n    this.maxSize = maxSize;\n  }\n\n  get(key: string): Float32Array | undefined {\n    const value = this.cache.get(key);\n    if (value) {\n      // Move to end (most recently used)\n      this.cache.delete(key);\n      this.cache.set(key, value);\n    }\n    return value;\n  }\n\n  set(key: string, value: Float32Array): void {\n    if (this.cache.size >= this.maxSize) {\n      // Delete oldest (first) entry\n      const firstKey = this.cache.keys().next().value;\n      if (firstKey) {\n        this.cache.delete(firstKey);\n      }\n    }\n    this.cache.set(key, value);\n  }\n\n  clear(): void {\n    this.cache.clear();\n  }\n\n  get size(): number {\n    return this.cache.size;\n  }\n}\n\nexport class EmbeddingService {\n  private static instance: EmbeddingService | null = null;\n\n  private backend: EmbeddingBackend;\n  private dimension: number;\n  private modelName: string;\n\n  // ONNX state\n  private embedder: WasmEmbedder | null = null;\n  private wasmModule: WasmModule | null = null;\n  private loadingPromise: Promise<void> | null = null;\n  private modelLoaded: boolean = false;\n\n  // Stats\n  private totalEmbeddings: number = 0;\n  private totalLatencyMs: number = 0;\n  private cacheHits: number = 0;\n\n  // Cache\n  private cache: EmbeddingCache;\n  private cacheEnabled: boolean;\n\n  private constructor() {\n    this.backend = (process.env.AGENTIC_FLOW_EMBEDDINGS as EmbeddingBackend) || 'simple';\n    this.modelName = process.env.AGENTIC_FLOW_EMBEDDING_MODEL || 'all-MiniLM-L6-v2';\n    this.dimension = this.backend === 'onnx' ? 384 : 256; // ONNX uses 384, simple uses 256\n    this.cacheEnabled = process.env.AGENTIC_FLOW_EMBEDDING_CACHE !== 'false';\n    this.cache = new EmbeddingCache(1000);\n  }\n\n  static getInstance(): EmbeddingService {\n    if (!EmbeddingService.instance) {\n      EmbeddingService.instance = new EmbeddingService();\n    }\n    return EmbeddingService.instance;\n  }\n\n  /**\n   * Get current backend\n   */\n  getBackend(): EmbeddingBackend {\n    return this.backend;\n  }\n\n  /**\n   * Get embedding dimension\n   */\n  getDimension(): number {\n    return this.dimension;\n  }\n\n  /**\n   * Check if ONNX model is loaded\n   */\n  isModelLoaded(): boolean {\n    return this.modelLoaded;\n  }\n\n  /**\n   * Initialize ONNX embedder (lazy load on first use)\n   */\n  async initOnnx(): Promise<void> {\n    if (this.modelLoaded) return;\n    if (this.loadingPromise) return this.loadingPromise;\n\n    this.loadingPromise = this._loadOnnxModel();\n    await this.loadingPromise;\n  }\n\n  private async _loadOnnxModel(): Promise<void> {\n    try {\n      console.log(`[EmbeddingService] Loading ONNX model: ${this.modelName}`);\n      const startTime = performance.now();\n\n      // Import WASM module (auto-initializes with --experimental-wasm-modules)\n      const wasmModule = await import('ruvector-onnx-embeddings-wasm');\n      this.wasmModule = wasmModule as unknown as WasmModule;\n\n      // Import loader for model fetching\n      const loaderModule = await import('ruvector-onnx-embeddings-wasm/loader.js');\n      const { ModelLoader, MODELS } = loaderModule;\n\n      const modelConfig = MODELS[this.modelName] as ModelConfig;\n      if (!modelConfig) {\n        throw new Error(`Unknown model: ${this.modelName}`);\n      }\n\n      // Load model files\n      const loader = new ModelLoader();\n      const { modelBytes, tokenizerJson } = await loader.loadModel(this.modelName);\n\n      // Create embedder with configuration\n      const embedderConfig = new (wasmModule as any).WasmEmbedderConfig()\n        .setMaxLength(modelConfig.maxLength)\n        .setNormalize(true)\n        .setPooling(0); // Mean pooling\n\n      this.embedder = (wasmModule as any).WasmEmbedder.withConfig(\n        modelBytes,\n        tokenizerJson,\n        embedderConfig\n      );\n\n      this.dimension = this.embedder!.dimension();\n      this.modelLoaded = true;\n\n      const loadTime = performance.now() - startTime;\n      console.log(`[EmbeddingService] Model loaded in ${loadTime.toFixed(0)}ms, dim=${this.dimension}`);\n    } catch (error) {\n      console.error('[EmbeddingService] Failed to load ONNX model:', error);\n      console.log('[EmbeddingService] Falling back to simple embeddings');\n      this.backend = 'simple';\n      this.dimension = 256;\n    }\n  }\n\n  /**\n   * Generate embedding for text\n   */\n  async embed(text: string): Promise<Float32Array> {\n    const startTime = performance.now();\n\n    // Check cache\n    if (this.cacheEnabled) {\n      const cached = this.cache.get(text);\n      if (cached) {\n        this.cacheHits++;\n        return cached;\n      }\n    }\n\n    let embedding: Float32Array;\n\n    if (this.backend === 'onnx') {\n      await this.initOnnx();\n      if (this.embedder) {\n        embedding = this.embedder.embedOne(text);\n      } else {\n        embedding = this.simpleEmbed(text);\n      }\n    } else {\n      embedding = this.simpleEmbed(text);\n    }\n\n    // Update stats\n    this.totalEmbeddings++;\n    this.totalLatencyMs += performance.now() - startTime;\n\n    // Cache result\n    if (this.cacheEnabled) {\n      this.cache.set(text, embedding);\n    }\n\n    return embedding;\n  }\n\n  /**\n   * Generate embeddings for multiple texts\n   */\n  async embedBatch(texts: string[]): Promise<Float32Array[]> {\n    if (this.backend === 'onnx' && this.embedder) {\n      await this.initOnnx();\n      if (this.embedder) {\n        const batchEmbedding = this.embedder.embedBatch(texts);\n        const dim = this.dimension;\n        const embeddings: Float32Array[] = [];\n        for (let i = 0; i < texts.length; i++) {\n          embeddings.push(batchEmbedding.slice(i * dim, (i + 1) * dim));\n        }\n        return embeddings;\n      }\n    }\n\n    // Fall back to sequential\n    return Promise.all(texts.map(t => this.embed(t)));\n  }\n\n  /**\n   * Compute similarity between two texts\n   */\n  async similarity(text1: string, text2: string): Promise<number> {\n    if (this.backend === 'onnx') {\n      await this.initOnnx();\n      if (this.embedder) {\n        return this.embedder.similarity(text1, text2);\n      }\n    }\n\n    // Fall back to embedding + cosine\n    const [e1, e2] = await Promise.all([this.embed(text1), this.embed(text2)]);\n    return this.cosineSimilarity(e1, e2);\n  }\n\n  /**\n   * Simple hash-based embedding (fast, not semantic)\n   */\n  simpleEmbed(text: string, dim: number = 256): Float32Array {\n    const embedding = new Float32Array(dim);\n\n    // Multi-pass hash for better distribution\n    for (let i = 0; i < text.length; i++) {\n      const code = text.charCodeAt(i);\n      embedding[i % dim] += code / 255;\n      embedding[(i * 7) % dim] += (code * 0.3) / 255;\n      embedding[(i * 13) % dim] += (code * 0.2) / 255;\n    }\n\n    // Normalize\n    let norm = 0;\n    for (let i = 0; i < dim; i++) {\n      norm += embedding[i] * embedding[i];\n    }\n    norm = Math.sqrt(norm) || 1;\n    for (let i = 0; i < dim; i++) {\n      embedding[i] /= norm;\n    }\n\n    return embedding;\n  }\n\n  /**\n   * Compute cosine similarity between two embeddings\n   */\n  cosineSimilarity(a: Float32Array, b: Float32Array): number {\n    if (this.wasmModule) {\n      return this.wasmModule.cosineSimilarity(a, b);\n    }\n\n    // JS fallback\n    let dot = 0;\n    let normA = 0;\n    let normB = 0;\n    for (let i = 0; i < a.length; i++) {\n      dot += a[i] * b[i];\n      normA += a[i] * a[i];\n      normB += b[i] * b[i];\n    }\n    return dot / (Math.sqrt(normA) * Math.sqrt(normB) || 1);\n  }\n\n  /**\n   * Get statistics\n   */\n  getStats(): EmbeddingStats {\n    return {\n      backend: this.backend,\n      dimension: this.dimension,\n      totalEmbeddings: this.totalEmbeddings,\n      totalLatencyMs: this.totalLatencyMs,\n      avgLatencyMs: this.totalEmbeddings > 0 ? this.totalLatencyMs / this.totalEmbeddings : 0,\n      cacheHits: this.cacheHits,\n      modelLoaded: this.modelLoaded,\n      modelName: this.backend === 'onnx' ? this.modelName : undefined,\n      simdAvailable: this.wasmModule?.simd_available?.() ?? undefined,\n    };\n  }\n\n  /**\n   * Clear cache\n   */\n  clearCache(): void {\n    this.cache.clear();\n  }\n\n  /**\n   * Reset instance (for testing)\n   */\n  static reset(): void {\n    if (EmbeddingService.instance?.embedder) {\n      EmbeddingService.instance.embedder.free();\n    }\n    EmbeddingService.instance = null;\n  }\n}\n\n// Export singleton getter\nexport function getEmbeddingService(): EmbeddingService {\n  return EmbeddingService.getInstance();\n}\n\n// Export convenience functions\nexport async function embed(text: string): Promise<Float32Array> {\n  return getEmbeddingService().embed(text);\n}\n\nexport async function embedBatch(texts: string[]): Promise<Float32Array[]> {\n  return getEmbeddingService().embedBatch(texts);\n}\n\nexport async function textSimilarity(text1: string, text2: string): Promise<number> {\n  return getEmbeddingService().similarity(text1, text2);\n}\n\nexport function simpleEmbed(text: string, dim: number = 256): Float32Array {\n  return getEmbeddingService().simpleEmbed(text, dim);\n}\n"]}
+{"version":3,"file":"EmbeddingService.js","sourceRoot":"","sources":["../../src/intelligence/EmbeddingService.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AA2DH,0BAA0B;AAC1B,IAAI,aAAa,GAAmB,IAAI,CAAC;AACzC,IAAI,cAAc,GAA0B,IAAI,CAAC;AAEjD;;GAEG;AACH,KAAK,UAAU,UAAU;IACvB,IAAI,aAAa,KAAK,IAAI,EAAE,CAAC;QAC3B,OAAO,aAAa,CAAC;IACvB,CAAC;IAED,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,MAAM,MAAM,CAAC,UAAU,CAA8B,CAAC;QAClE,cAAc,GAAG,GAAG,CAAC;QACrB,aAAa,GAAG,GAAG,CAAC,eAAe,EAAE,EAAE,IAAI,KAAK,CAAC;QACjD,OAAO,aAAa,CAAC;IACvB,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,2DAA2D;QAC3D,aAAa,GAAG,KAAK,CAAC;QACtB,OAAO,KAAK,CAAC;IACf,CAAC;AACH,CAAC;AAED,kCAAkC;AAClC,MAAM,cAAc;IACV,KAAK,GAA8B,IAAI,GAAG,EAAE,CAAC;IAC7C,OAAO,CAAS;IAExB,YAAY,UAAkB,IAAI;QAChC,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;IACzB,CAAC;IAED,GAAG,CAAC,GAAW;QACb,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAClC,IAAI,KAAK,EAAE,CAAC;YACV,mCAAmC;YACnC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;YACvB,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;QAC7B,CAAC;QACD,OAAO,KAAK,CAAC;IACf,CAAC;IAED,GAAG,CAAC,GAAW,EAAE,KAAmB;QAClC,IAAI,IAAI,CAAC,KAAK,CAAC,IAAI,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACpC,8BAA8B;YAC9B,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC;YAChD,IAAI,QAAQ,EAAE,CAAC;gBACb,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;YAC9B,CAAC;QACH,CAAC;QACD,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;IAC7B,CAAC;IAED,KAAK;QACH,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,CAAC;IACrB,CAAC;IAED,IAAI,IAAI;QACN,OAAO,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC;IACzB,CAAC;CACF;AAED,MAAM,OAAO,gBAAgB;IACnB,MAAM,CAAC,QAAQ,GAA4B,IAAI,CAAC;IAEhD,OAAO,CAAmB;IAC1B,gBAAgB,GAA4B,IAAI,CAAC;IACjD,SAAS,CAAS;IAClB,SAAS,CAAS;IAE1B,aAAa;IACL,WAAW,GAAY,KAAK,CAAC;IAC7B,cAAc,GAAyB,IAAI,CAAC;IAEpD,QAAQ;IACA,eAAe,GAAW,CAAC,CAAC;IAC5B,cAAc,GAAW,CAAC,CAAC;IAC3B,SAAS,GAAW,CAAC,CAAC;IAE9B,QAAQ;IACA,KAAK,CAAiB;IACtB,YAAY,CAAU;IAE9B,+BAA+B;IACvB,MAAM,GAAoD,EAAE,KAAK,EAAE,EAAE,EAAE,UAAU,EAAE,EAAE,EAAE,CAAC;IAEhG;QACE,mEAAmE;QACnE,IAAI,CAAC,OAAO,GAAI,OAAO,CAAC,GAAG,CAAC,uBAA4C,IAAI,MAAM,CAAC;QACnF,IAAI,CAAC,SAAS,GAAG,OAAO,CAAC,GAAG,CAAC,4BAA4B,IAAI,kBAAkB,CAAC;QAChF,IAAI,CAAC,SAAS,GAAG,GAAG,CAAC,CAAC,wCAAwC;QAC9D,IAAI,CAAC,YAAY,GAAG,OAAO,CAAC,GAAG,CAAC,4BAA4B,KAAK,OAAO,CAAC;QACzE,IAAI,CAAC,KAAK,GAAG,IAAI,cAAc,CAAC,IAAI,CAAC,CAAC;IACxC,CAAC;IAED,MAAM,CAAC,WAAW;QAChB,IAAI,CAAC,gBAAgB,CAAC,QAAQ,EAAE,CAAC;YAC/B,gBAAgB,CAAC,QAAQ,GAAG,IAAI,gBAAgB,EAAE,CAAC;QACrD,CAAC;QACD,OAAO,gBAAgB,CAAC,QAAQ,CAAC;IACnC,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,cAAc;QAC1B,IAAI,IAAI,CAAC,gBAAgB,EAAE,CAAC;YAC1B,OAAO,IAAI,CAAC,gBAAgB,CAAC;QAC/B,CAAC;QAED,IAAI,IAAI,CAAC,OAAO,KAAK,MAAM,EAAE,CAAC;YAC5B,MAAM,OAAO,GAAG,MAAM,UAAU,EAAE,CAAC;YACnC,IAAI,CAAC,gBAAgB,GAAG,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC;YACpD,IAAI,OAAO,EAAE,CAAC;gBACZ,IAAI,CAAC,SAAS,GAAG,GAAG,CAAC,CAAC,6BAA6B;YACrD,CAAC;QACH,CAAC;aAAM,CAAC;YACN,IAAI,CAAC,gBAAgB,GAAG,IAAI,CAAC,OAAO,CAAC;YACrC,IAAI,IAAI,CAAC,OAAO,KAAK,MAAM,EAAE,CAAC;gBAC5B,MAAM,UAAU,EAAE,CAAC,CAAC,0BAA0B;gBAC9C,IAAI,CAAC,SAAS,GAAG,GAAG,CAAC;YACvB,CAAC;QACH,CAAC;QAED,OAAO,IAAI,CAAC,gBAAgB,CAAC;IAC/B,CAAC;IAED;;OAEG;IACH,UAAU;QACR,OAAO,IAAI,CAAC,OAAO,CAAC;IACtB,CAAC;IAED;;OAEG;IACH,mBAAmB;QACjB,OAAO,IAAI,CAAC,gBAAgB,IAAI,IAAI,CAAC,OAAO,CAAC;IAC/C,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO,IAAI,CAAC,SAAS,CAAC;IACxB,CAAC;IAED;;OAEG;IACH,aAAa;QACX,OAAO,IAAI,CAAC,WAAW,CAAC;IAC1B,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,KAAK,CAAC,IAAY;QACtB,MAAM,SAAS,GAAG,WAAW,CAAC,GAAG,EAAE,CAAC;QAEpC,cAAc;QACd,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;YACtB,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;YACpC,IAAI,MAAM,EAAE,CAAC;gBACX,IAAI,CAAC,SAAS,EAAE,CAAC;gBACjB,OAAO,MAAM,CAAC;YAChB,CAAC;QACH,CAAC;QAED,wCAAwC;QACxC,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;QACrD,IAAI,SAAuB,CAAC;QAE5B,IAAI,gBAAgB,KAAK,MAAM,IAAI,cAAc,EAAE,CAAC;YAClD,MAAM,MAAM,GAAG,MAAM,cAAc,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;YAChD,IAAI,MAAM,EAAE,SAAS,EAAE,CAAC;gBACtB,SAAS,GAAG,MAAM,CAAC,SAAS,CAAC;gBAC7B,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC;YAC1B,CAAC;iBAAM,CAAC;gBACN,SAAS,GAAG,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;YACrC,CAAC;QACH,CAAC;aAAM,CAAC;YACN,SAAS,GAAG,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,CAAC;QACrC,CAAC;QAED,eAAe;QACf,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,IAAI,CAAC,cAAc,IAAI,WAAW,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;QAErD,eAAe;QACf,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;YACtB,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,SAAS,CAAC,CAAC;QAClC,CAAC;QAED,OAAO,SAAS,CAAC;IACnB,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,UAAU,CAAC,KAAe;QAC9B,MAAM,SAAS,GAAG,WAAW,CAAC,GAAG,EAAE,CAAC;QAEpC,kCAAkC;QAClC,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;YACtB,MAAM,aAAa,GAA4B,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC;YACzF,MAAM,SAAS,GAAG,aAAa,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,KAAK,IAAI,CAAC,CAAC;YACvD,IAAI,SAAS,EAAE,CAAC;gBACd,IAAI,CAAC,SAAS,IAAI,KAAK,CAAC,MAAM,CAAC;gBAC/B,OAAO,aAA+B,CAAC;YACzC,CAAC;QACH,CAAC;QAED,kBAAkB;QAClB,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;QAErD,IAAI,gBAAgB,KAAK,MAAM,IAAI,cAAc,EAAE,CAAC;YAClD,MAAM,MAAM,GAAG,MAAM,cAAc,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;YACtD,IAAI,MAAM,EAAE,UAAU,IAAI,MAAM,CAAC,UAAU,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,EAAE,CAAC;gBACpE,MAAM,UAAU,GAAG,MAAM,CAAC,UAAU,CAAC;gBAErC,8BAA8B;gBAC9B,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;oBACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;wBACtC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;oBAC1C,CAAC;gBACH,CAAC;gBAED,eAAe;gBACf,IAAI,CAAC,eAAe,IAAI,KAAK,CAAC,MAAM,CAAC;gBACrC,IAAI,CAAC,cAAc,IAAI,WAAW,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;gBACrD,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC;gBAExB,OAAO,UAAU,CAAC;YACpB,CAAC;QACH,CAAC;QAED,6CAA6C;QAC7C,OAAO,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,UAAU,CAAC,KAAa,EAAE,KAAa;QAC3C,MAAM,gBAAgB,GAAG,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;QAErD,IAAI,gBAAgB,KAAK,MAAM,IAAI,cAAc,EAAE,CAAC;YAClD,MAAM,MAAM,GAAG,MAAM,cAAc,CAAC,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;YAC7D,OAAO,MAAM,CAAC,UAAU,CAAC;QAC3B,CAAC;QAED,kCAAkC;QAClC,MAAM,CAAC,EAAE,EAAE,EAAE,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;QAC3E,OAAO,IAAI,CAAC,gBAAgB,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC;IACvC,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,gBAAgB,CAAC,KAAe;QACpC,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;QAChD,MAAM,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC;QACvB,MAAM,MAAM,GAAe,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC;QAE3E,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,kBAAkB;YACtC,KAAK,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC/B,MAAM,GAAG,GAAG,IAAI,CAAC,gBAAgB,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;gBAChE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC;gBACnB,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,YAAY;YAClC,CAAC;QACH,CAAC;QAED,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,WAAW,CAAC,KAAe;QAC/B,IAAI,CAAC,MAAM,CAAC,KAAK,GAAG,KAAK,CAAC;QAC1B,IAAI,CAAC,MAAM,CAAC,UAAU,GAAG,MAAM,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;IACxD,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,cAAc,CAAC,KAAa,EAAE,OAAe,CAAC;QAClD,IAAI,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,KAAK,CAAC,6CAA6C,CAAC,CAAC;QACjE,CAAC;QAED,MAAM,cAAc,GAAG,MAAM,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;QAC/C,MAAM,OAAO,GAAmB,EAAE,CAAC;QAEnC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAClD,MAAM,GAAG,GAAG,IAAI,CAAC,gBAAgB,CAAC,cAAc,EAAE,IAAI,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;YAC7E,OAAO,CAAC,IAAI,CAAC;gBACX,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;gBAC1B,KAAK,EAAE,CAAC;gBACR,UAAU,EAAE,GAAG;aAChB,CAAC,CAAC;QACL,CAAC;QAED,mDAAmD;QACnD,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,GAAG,CAAC,CAAC,UAAU,CAAC,CAAC;QACpD,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;IAChC,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,cAAc,CAAC,KAAe,EAAE,YAAoB,GAAG;QAC3D,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;QAChD,MAAM,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC;QACvB,MAAM,OAAO,GAAG,IAAI,GAAG,EAAU,CAAC;QAClC,MAAM,MAAM,GAAqB,EAAE,CAAC;QAEpC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,IAAI,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;gBAAE,SAAS;YAE7B,MAAM,KAAK,GAAmB;gBAC5B,OAAO,EAAE,CAAC,CAAC,CAAC;gBACZ,KAAK,EAAE,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;gBACjB,UAAU,EAAE,GAAG;aAChB,CAAC;YAEF,KAAK,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC/B,IAAI,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;oBAAE,SAAS;gBAE7B,MAAM,GAAG,GAAG,IAAI,CAAC,gBAAgB,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;gBAChE,IAAI,GAAG,IAAI,SAAS,EAAE,CAAC;oBACrB,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;oBACtB,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;oBAC3B,KAAK,CAAC,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,UAAU,EAAE,GAAG,CAAC,CAAC;oBACnD,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;gBACjB,CAAC;YACH,CAAC;YAED,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAC7B,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;gBACf,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YACrB,CAAC;QACH,CAAC;QAED,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,YAAY,CAChB,KAAe,EACf,IAAY,CAAC,EACb,gBAAwB,GAAG;QAE3B,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;QAChD,MAAM,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC;QACvB,MAAM,GAAG,GAAG,IAAI,CAAC,SAAS,CAAC;QAE3B,gFAAgF;QAChF,MAAM,eAAe,GAAG,IAAI,GAAG,EAAU,CAAC;QAC1C,OAAO,eAAe,CAAC,IAAI,GAAG,CAAC,IAAI,eAAe,CAAC,IAAI,GAAG,CAAC,EAAE,CAAC;YAC5D,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC;QACrD,CAAC;QACD,IAAI,SAAS,GAAmB,KAAK,CAAC,IAAI,CAAC,eAAe,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE;YAClE,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC;YACnC,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;YACxB,OAAO,IAAI,CAAC;QACd,CAAC,CAAC,CAAC;QAEH,IAAI,QAAQ,GAAG,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAEpC,KAAK,IAAI,IAAI,GAAG,CAAC,EAAE,IAAI,GAAG,aAAa,EAAE,IAAI,EAAE,EAAE,CAAC;YAChD,oCAAoC;YACpC,MAAM,WAAW,GAAG,UAAU,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE;gBACvC,IAAI,WAAW,GAAG,CAAC,CAAC;gBACpB,IAAI,OAAO,GAAG,CAAC,QAAQ,CAAC;gBACxB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;oBAC3B,MAAM,GAAG,GAAG,IAAI,CAAC,gBAAgB,CAAC,GAAG,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;oBACrD,IAAI,GAAG,GAAG,OAAO,EAAE,CAAC;wBAClB,OAAO,GAAG,GAAG,CAAC;wBACd,WAAW,GAAG,CAAC,CAAC;oBAClB,CAAC;gBACH,CAAC;gBACD,OAAO,WAAW,CAAC;YACrB,CAAC,CAAC,CAAC;YAEH,oBAAoB;YACpB,MAAM,OAAO,GAAG,WAAW,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;YAC9D,QAAQ,GAAG,WAAW,CAAC;YACvB,IAAI,CAAC,OAAO;gBAAE,MAAM;YAEpB,mBAAmB;YACnB,MAAM,YAAY,GAAmB,EAAE,CAAC;YACxC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,YAAY,CAAC,IAAI,CAAC,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC;YAC3C,CAAC;YACD,MAAM,MAAM,GAAG,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;YAEpC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,MAAM,CAAC,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;gBACtB,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC;gBACZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;oBAC7B,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;gBACzC,CAAC;YACH,CAAC;YAED,sBAAsB;YACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,IAAI,MAAM,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC;oBAClB,IAAI,IAAI,GAAG,CAAC,CAAC;oBACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;wBAC7B,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,MAAM,CAAC,CAAC,CAAC,CAAC;wBAChC,IAAI,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;oBAClD,CAAC;oBACD,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;oBAC5B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;wBAC7B,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;oBAC7B,CAAC;gBACH,CAAC;YACH,CAAC;YACD,SAAS,GAAG,YAAY,CAAC;QAC3B,CAAC;QAED,OAAO,EAAE,QAAQ,EAAE,SAAS,EAAE,CAAC;IACjC,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,CAAC,WAAW,CAAC,KAAe,EAAE,YAAoB,EAAE;QACxD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,IAAI,SAAS,EAAE,CAAC;YACjD,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,GAAG,SAAS,CAAC,CAAC;YAC5C,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;YAEhD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACtC,MAAM;oBACJ,KAAK,EAAE,CAAC,GAAG,CAAC;oBACZ,IAAI,EAAE,KAAK,CAAC,CAAC,CAAC;oBACd,SAAS,EAAE,UAAU,CAAC,CAAC,CAAC;iBACzB,CAAC;YACJ,CAAC;QACH,CAAC;IACH,CAAC;IAED;;OAEG;IACH,WAAW,CAAC,IAAY,EAAE,MAAc,GAAG;QACzC,MAAM,SAAS,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,CAAC;QAExC,0CAA0C;QAC1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACrC,MAAM,IAAI,GAAG,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC;YAChC,SAAS,CAAC,CAAC,GAAG,GAAG,CAAC,IAAI,IAAI,GAAG,GAAG,CAAC;YACjC,SAAS,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC,IAAI,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC;YAC/C,SAAS,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,GAAG,GAAG,CAAC,IAAI,CAAC,IAAI,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC;QAClD,CAAC;QAED,YAAY;QACZ,IAAI,IAAI,GAAG,CAAC,CAAC;QACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7B,IAAI,IAAI,SAAS,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC;QACtC,CAAC;QACD,IAAI,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QAC5B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7B,SAAS,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;QACvB,CAAC;QAED,OAAO,SAAS,CAAC;IACnB,CAAC;IAED;;OAEG;IACH,gBAAgB,CAAC,CAAe,EAAE,CAAe;QAC/C,IAAI,cAAc,EAAE,gBAAgB,EAAE,CAAC;YACrC,OAAO,cAAc,CAAC,gBAAgB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QAC/C,CAAC;QAED,cAAc;QACd,IAAI,GAAG,GAAG,CAAC,CAAC;QACZ,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAClC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACnB,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;QACvB,CAAC;QACD,OAAO,GAAG,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC;IAC1D,CAAC;IAED;;OAEG;IACH,QAAQ;QACN,MAAM,SAAS,GAAG,IAAI,CAAC,gBAAgB,IAAI,IAAI,CAAC,OAAO,CAAC;QACxD,MAAM,aAAa,GAAG,cAAc,EAAE,QAAQ,EAAE,EAAE,IAAI,EAAE,CAAC;QAEzD,OAAO;YACL,OAAO,EAAE,IAAI,CAAC,OAAO;YACrB,gBAAgB,EAAE,SAAS;YAC3B,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,eAAe,EAAE,IAAI,CAAC,eAAe;YACrC,cAAc,EAAE,IAAI,CAAC,cAAc;YACnC,YAAY,EAAE,IAAI,CAAC,eAAe,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,GAAG,IAAI,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;YACvF,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,WAAW,EAAE,IAAI,CAAC,WAAW;YAC7B,SAAS,EAAE,SAAS,KAAK,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS;YAC5D,aAAa,EAAE,aAAa,CAAC,aAAa,IAAI,aAAa;YAC3D,eAAe,EAAE,aAAa,CAAC,WAAW,IAAI,SAAS;SACxD,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,UAAU;QACR,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,CAAC;IACrB,CAAC;IAED;;OAEG;IACH,WAAW;QACT,IAAI,CAAC,MAAM,GAAG,EAAE,KAAK,EAAE,EAAE,EAAE,UAAU,EAAE,EAAE,EAAE,CAAC;IAC9C,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,QAAQ;QACZ,IAAI,cAAc,EAAE,QAAQ,EAAE,CAAC;YAC7B,MAAM,cAAc,CAAC,QAAQ,EAAE,CAAC;QAClC,CAAC;IACH,CAAC;IAED;;OAEG;IACH,MAAM,CAAC,KAAK,CAAC,KAAK;QAChB,IAAI,gBAAgB,CAAC,QAAQ,EAAE,CAAC;YAC9B,MAAM,gBAAgB,CAAC,QAAQ,CAAC,QAAQ,EAAE,CAAC;QAC7C,CAAC;QACD,gBAAgB,CAAC,QAAQ,GAAG,IAAI,CAAC;QACjC,aAAa,GAAG,IAAI,CAAC;QACrB,cAAc,GAAG,IAAI,CAAC;IACxB,CAAC;;AAGH,0BAA0B;AAC1B,MAAM,UAAU,mBAAmB;IACjC,OAAO,gBAAgB,CAAC,WAAW,EAAE,CAAC;AACxC,CAAC;AAED,+BAA+B;AAC/B,MAAM,CAAC,KAAK,UAAU,KAAK,CAAC,IAAY;IACtC,OAAO,mBAAmB,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;AAC3C,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,UAAU,CAAC,KAAe;IAC9C,OAAO,mBAAmB,EAAE,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;AACjD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,KAAa,EAAE,KAAa;IAC/D,OAAO,mBAAmB,EAAE,CAAC,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;AACxD,CAAC;AAED,MAAM,UAAU,WAAW,CAAC,IAAY,EAAE,MAAc,GAAG;IACzD,OAAO,mBAAmB,EAAE,CAAC,WAAW,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;AACtD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,KAAe;IACpD,OAAO,mBAAmB,EAAE,CAAC,gBAAgB,CAAC,KAAK,CAAC,CAAC;AACvD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,KAAa,EAAE,OAAe,CAAC;IAClE,OAAO,mBAAmB,EAAE,CAAC,cAAc,CAAC,KAAK,EAAE,IAAI,CAAC,CAAC;AAC3D,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,KAAe,EAAE,YAAoB,GAAG;IAC3E,OAAO,mBAAmB,EAAE,CAAC,cAAc,CAAC,KAAK,EAAE,SAAS,CAAC,CAAC;AAChE,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,YAAY,CAAC,KAAe,EAAE,IAAY,CAAC;IAC/D,OAAO,mBAAmB,EAAE,CAAC,YAAY,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;AACtD,CAAC","sourcesContent":["/**\n * EmbeddingService - Unified embedding interface for agentic-flow\n *\n * Uses ruvector@0.1.61+ for ONNX embeddings with:\n * - SIMD128 acceleration (6x faster)\n * - Parallel worker threads (7 workers)\n * - all-MiniLM-L6-v2 model (384 dimensions)\n *\n * Configure via:\n * - AGENTIC_FLOW_EMBEDDINGS=simple|onnx|auto (default: auto)\n * - AGENTIC_FLOW_EMBEDDING_MODEL=all-MiniLM-L6-v2 (default)\n */\n\nexport type EmbeddingBackend = 'simple' | 'onnx' | 'auto';\n\nexport interface EmbeddingStats {\n  backend: EmbeddingBackend;\n  effectiveBackend: EmbeddingBackend;\n  dimension: number;\n  totalEmbeddings: number;\n  totalLatencyMs: number;\n  avgLatencyMs: number;\n  cacheHits: number;\n  modelLoaded: boolean;\n  modelName?: string;\n  simdAvailable?: boolean;\n  parallelWorkers?: number;\n}\n\nexport interface SimilarityResult {\n  similarity: number;\n  timeMs: number;\n}\n\nexport interface SearchResult {\n  text: string;\n  index: number;\n  similarity: number;\n}\n\nexport interface DuplicateGroup {\n  indices: number[];\n  texts: string[];\n  similarity: number;\n}\n\n// Ruvector embedding result types\ninterface EmbeddingResult {\n  embedding: Float32Array;\n  timeMs?: number;\n}\n\ninterface BatchEmbeddingResult {\n  embeddings: Float32Array[];\n  timeMs?: number;\n}\n\n// Ruvector module interface\ninterface RuvectorModule {\n  isOnnxAvailable: () => boolean;\n  getDefaultEmbeddingService: () => any;\n  embed: (text: string) => Promise<EmbeddingResult | null>;\n  embedBatch: (texts: string[]) => Promise<BatchEmbeddingResult | null>;\n  similarity: (text1: string, text2: string) => Promise<SimilarityResult>;\n  toFloat32Array: (arr: number[]) => Float32Array;\n  cosineSimilarity: (a: Float32Array, b: Float32Array) => number;\n  getStats: () => any;\n  shutdown: () => Promise<void>;\n}\n\n// ONNX availability cache\nlet onnxAvailable: boolean | null = null;\nlet ruvectorModule: RuvectorModule | null = null;\n\n/**\n * Detect ONNX/SIMD support by loading ruvector\n */\nasync function detectOnnx(): Promise<boolean> {\n  if (onnxAvailable !== null) {\n    return onnxAvailable;\n  }\n\n  try {\n    const mod = await import('ruvector') as unknown as RuvectorModule;\n    ruvectorModule = mod;\n    onnxAvailable = mod.isOnnxAvailable?.() ?? false;\n    return onnxAvailable;\n  } catch (error) {\n    // Ruvector loading failed - fall back to simple embeddings\n    onnxAvailable = false;\n    return false;\n  }\n}\n\n// Simple LRU cache for embeddings\nclass EmbeddingCache {\n  private cache: Map<string, Float32Array> = new Map();\n  private maxSize: number;\n\n  constructor(maxSize: number = 1000) {\n    this.maxSize = maxSize;\n  }\n\n  get(key: string): Float32Array | undefined {\n    const value = this.cache.get(key);\n    if (value) {\n      // Move to end (most recently used)\n      this.cache.delete(key);\n      this.cache.set(key, value);\n    }\n    return value;\n  }\n\n  set(key: string, value: Float32Array): void {\n    if (this.cache.size >= this.maxSize) {\n      // Delete oldest (first) entry\n      const firstKey = this.cache.keys().next().value;\n      if (firstKey) {\n        this.cache.delete(firstKey);\n      }\n    }\n    this.cache.set(key, value);\n  }\n\n  clear(): void {\n    this.cache.clear();\n  }\n\n  get size(): number {\n    return this.cache.size;\n  }\n}\n\nexport class EmbeddingService {\n  private static instance: EmbeddingService | null = null;\n\n  private backend: EmbeddingBackend;\n  private effectiveBackend: EmbeddingBackend | null = null;\n  private dimension: number;\n  private modelName: string;\n\n  // ONNX state\n  private modelLoaded: boolean = false;\n  private loadingPromise: Promise<void> | null = null;\n\n  // Stats\n  private totalEmbeddings: number = 0;\n  private totalLatencyMs: number = 0;\n  private cacheHits: number = 0;\n\n  // Cache\n  private cache: EmbeddingCache;\n  private cacheEnabled: boolean;\n\n  // Corpus for search operations\n  private corpus: { texts: string[]; embeddings: Float32Array[] } = { texts: [], embeddings: [] };\n\n  private constructor() {\n    // Default to 'auto' which will detect ONNX and use it if available\n    this.backend = (process.env.AGENTIC_FLOW_EMBEDDINGS as EmbeddingBackend) || 'auto';\n    this.modelName = process.env.AGENTIC_FLOW_EMBEDDING_MODEL || 'all-MiniLM-L6-v2';\n    this.dimension = 256; // Will be updated when ONNX loads (384)\n    this.cacheEnabled = process.env.AGENTIC_FLOW_EMBEDDING_CACHE !== 'false';\n    this.cache = new EmbeddingCache(1000);\n  }\n\n  static getInstance(): EmbeddingService {\n    if (!EmbeddingService.instance) {\n      EmbeddingService.instance = new EmbeddingService();\n    }\n    return EmbeddingService.instance;\n  }\n\n  /**\n   * Resolve the effective backend based on ONNX detection\n   */\n  private async resolveBackend(): Promise<EmbeddingBackend> {\n    if (this.effectiveBackend) {\n      return this.effectiveBackend;\n    }\n\n    if (this.backend === 'auto') {\n      const hasOnnx = await detectOnnx();\n      this.effectiveBackend = hasOnnx ? 'onnx' : 'simple';\n      if (hasOnnx) {\n        this.dimension = 384; // all-MiniLM-L6-v2 dimension\n      }\n    } else {\n      this.effectiveBackend = this.backend;\n      if (this.backend === 'onnx') {\n        await detectOnnx(); // Ensure module is loaded\n        this.dimension = 384;\n      }\n    }\n\n    return this.effectiveBackend;\n  }\n\n  /**\n   * Get configured backend (may be 'auto')\n   */\n  getBackend(): EmbeddingBackend {\n    return this.backend;\n  }\n\n  /**\n   * Get effective backend after detection\n   */\n  getEffectiveBackend(): EmbeddingBackend {\n    return this.effectiveBackend || this.backend;\n  }\n\n  /**\n   * Get embedding dimension\n   */\n  getDimension(): number {\n    return this.dimension;\n  }\n\n  /**\n   * Check if ONNX model is loaded\n   */\n  isModelLoaded(): boolean {\n    return this.modelLoaded;\n  }\n\n  /**\n   * Generate embedding for text\n   * Auto-detects ONNX and uses it if available (default behavior)\n   */\n  async embed(text: string): Promise<Float32Array> {\n    const startTime = performance.now();\n\n    // Check cache\n    if (this.cacheEnabled) {\n      const cached = this.cache.get(text);\n      if (cached) {\n        this.cacheHits++;\n        return cached;\n      }\n    }\n\n    // Resolve backend (handles 'auto' mode)\n    const effectiveBackend = await this.resolveBackend();\n    let embedding: Float32Array;\n\n    if (effectiveBackend === 'onnx' && ruvectorModule) {\n      const result = await ruvectorModule.embed(text);\n      if (result?.embedding) {\n        embedding = result.embedding;\n        this.modelLoaded = true;\n      } else {\n        embedding = this.simpleEmbed(text);\n      }\n    } else {\n      embedding = this.simpleEmbed(text);\n    }\n\n    // Update stats\n    this.totalEmbeddings++;\n    this.totalLatencyMs += performance.now() - startTime;\n\n    // Cache result\n    if (this.cacheEnabled) {\n      this.cache.set(text, embedding);\n    }\n\n    return embedding;\n  }\n\n  /**\n   * Generate embeddings for multiple texts (batch processing with parallel workers)\n   * Batch processing provides significant speedup with parallel ONNX workers\n   */\n  async embedBatch(texts: string[]): Promise<Float32Array[]> {\n    const startTime = performance.now();\n\n    // Check cache for all texts first\n    if (this.cacheEnabled) {\n      const cachedResults: (Float32Array | null)[] = texts.map(t => this.cache.get(t) || null);\n      const allCached = cachedResults.every(r => r !== null);\n      if (allCached) {\n        this.cacheHits += texts.length;\n        return cachedResults as Float32Array[];\n      }\n    }\n\n    // Resolve backend\n    const effectiveBackend = await this.resolveBackend();\n\n    if (effectiveBackend === 'onnx' && ruvectorModule) {\n      const result = await ruvectorModule.embedBatch(texts);\n      if (result?.embeddings && result.embeddings.length === texts.length) {\n        const embeddings = result.embeddings;\n\n        // Cache individual embeddings\n        if (this.cacheEnabled) {\n          for (let i = 0; i < texts.length; i++) {\n            this.cache.set(texts[i], embeddings[i]);\n          }\n        }\n\n        // Update stats\n        this.totalEmbeddings += texts.length;\n        this.totalLatencyMs += performance.now() - startTime;\n        this.modelLoaded = true;\n\n        return embeddings;\n      }\n    }\n\n    // Fall back to sequential for simple backend\n    return Promise.all(texts.map(t => this.embed(t)));\n  }\n\n  /**\n   * Compute similarity between two texts\n   */\n  async similarity(text1: string, text2: string): Promise<number> {\n    const effectiveBackend = await this.resolveBackend();\n\n    if (effectiveBackend === 'onnx' && ruvectorModule) {\n      const result = await ruvectorModule.similarity(text1, text2);\n      return result.similarity;\n    }\n\n    // Fall back to embedding + cosine\n    const [e1, e2] = await Promise.all([this.embed(text1), this.embed(text2)]);\n    return this.cosineSimilarity(e1, e2);\n  }\n\n  /**\n   * Compute NxN similarity matrix for a list of texts\n   * Uses parallel workers for ONNX backend\n   */\n  async similarityMatrix(texts: string[]): Promise<number[][]> {\n    const embeddings = await this.embedBatch(texts);\n    const n = texts.length;\n    const matrix: number[][] = Array(n).fill(null).map(() => Array(n).fill(0));\n\n    for (let i = 0; i < n; i++) {\n      matrix[i][i] = 1.0; // Self-similarity\n      for (let j = i + 1; j < n; j++) {\n        const sim = this.cosineSimilarity(embeddings[i], embeddings[j]);\n        matrix[i][j] = sim;\n        matrix[j][i] = sim; // Symmetric\n      }\n    }\n\n    return matrix;\n  }\n\n  /**\n   * Build a corpus for semantic search\n   */\n  async buildCorpus(texts: string[]): Promise<void> {\n    this.corpus.texts = texts;\n    this.corpus.embeddings = await this.embedBatch(texts);\n  }\n\n  /**\n   * Semantic search against the corpus\n   * Returns top-k most similar texts\n   */\n  async semanticSearch(query: string, topK: number = 5): Promise<SearchResult[]> {\n    if (this.corpus.texts.length === 0) {\n      throw new Error('Corpus not built. Call buildCorpus() first.');\n    }\n\n    const queryEmbedding = await this.embed(query);\n    const results: SearchResult[] = [];\n\n    for (let i = 0; i < this.corpus.texts.length; i++) {\n      const sim = this.cosineSimilarity(queryEmbedding, this.corpus.embeddings[i]);\n      results.push({\n        text: this.corpus.texts[i],\n        index: i,\n        similarity: sim,\n      });\n    }\n\n    // Sort by similarity (descending) and return top-k\n    results.sort((a, b) => b.similarity - a.similarity);\n    return results.slice(0, topK);\n  }\n\n  /**\n   * Find near-duplicate texts in a list\n   * Groups texts with similarity above threshold\n   */\n  async findDuplicates(texts: string[], threshold: number = 0.9): Promise<DuplicateGroup[]> {\n    const embeddings = await this.embedBatch(texts);\n    const n = texts.length;\n    const visited = new Set<number>();\n    const groups: DuplicateGroup[] = [];\n\n    for (let i = 0; i < n; i++) {\n      if (visited.has(i)) continue;\n\n      const group: DuplicateGroup = {\n        indices: [i],\n        texts: [texts[i]],\n        similarity: 1.0,\n      };\n\n      for (let j = i + 1; j < n; j++) {\n        if (visited.has(j)) continue;\n\n        const sim = this.cosineSimilarity(embeddings[i], embeddings[j]);\n        if (sim >= threshold) {\n          group.indices.push(j);\n          group.texts.push(texts[j]);\n          group.similarity = Math.min(group.similarity, sim);\n          visited.add(j);\n        }\n      }\n\n      if (group.indices.length > 1) {\n        visited.add(i);\n        groups.push(group);\n      }\n    }\n\n    return groups;\n  }\n\n  /**\n   * K-means clustering of texts\n   * Returns cluster assignments and centroids\n   */\n  async clusterTexts(\n    texts: string[],\n    k: number = 3,\n    maxIterations: number = 100\n  ): Promise<{ clusters: number[]; centroids: Float32Array[] }> {\n    const embeddings = await this.embedBatch(texts);\n    const n = texts.length;\n    const dim = this.dimension;\n\n    // Initialize centroids randomly (copy to new ArrayBuffer for consistent typing)\n    const centroidIndices = new Set<number>();\n    while (centroidIndices.size < k && centroidIndices.size < n) {\n      centroidIndices.add(Math.floor(Math.random() * n));\n    }\n    let centroids: Float32Array[] = Array.from(centroidIndices).map(i => {\n      const copy = new Float32Array(dim);\n      copy.set(embeddings[i]);\n      return copy;\n    });\n\n    let clusters = new Array(n).fill(0);\n\n    for (let iter = 0; iter < maxIterations; iter++) {\n      // Assign points to nearest centroid\n      const newClusters = embeddings.map(emb => {\n        let bestCluster = 0;\n        let bestSim = -Infinity;\n        for (let c = 0; c < k; c++) {\n          const sim = this.cosineSimilarity(emb, centroids[c]);\n          if (sim > bestSim) {\n            bestSim = sim;\n            bestCluster = c;\n          }\n        }\n        return bestCluster;\n      });\n\n      // Check convergence\n      const changed = newClusters.some((c, i) => c !== clusters[i]);\n      clusters = newClusters;\n      if (!changed) break;\n\n      // Update centroids\n      const newCentroids: Float32Array[] = [];\n      for (let c = 0; c < k; c++) {\n        newCentroids.push(new Float32Array(dim));\n      }\n      const counts = new Array(k).fill(0);\n\n      for (let i = 0; i < n; i++) {\n        const c = clusters[i];\n        counts[c]++;\n        for (let d = 0; d < dim; d++) {\n          newCentroids[c][d] += embeddings[i][d];\n        }\n      }\n\n      // Normalize centroids\n      for (let c = 0; c < k; c++) {\n        if (counts[c] > 0) {\n          let norm = 0;\n          for (let d = 0; d < dim; d++) {\n            newCentroids[c][d] /= counts[c];\n            norm += newCentroids[c][d] * newCentroids[c][d];\n          }\n          norm = Math.sqrt(norm) || 1;\n          for (let d = 0; d < dim; d++) {\n            newCentroids[c][d] /= norm;\n          }\n        }\n      }\n      centroids = newCentroids;\n    }\n\n    return { clusters, centroids };\n  }\n\n  /**\n   * Stream embeddings for large batches (memory efficient)\n   * Yields embeddings one at a time\n   */\n  async *streamEmbed(texts: string[], batchSize: number = 32): AsyncGenerator<{ index: number; text: string; embedding: Float32Array }> {\n    for (let i = 0; i < texts.length; i += batchSize) {\n      const batch = texts.slice(i, i + batchSize);\n      const embeddings = await this.embedBatch(batch);\n\n      for (let j = 0; j < batch.length; j++) {\n        yield {\n          index: i + j,\n          text: batch[j],\n          embedding: embeddings[j],\n        };\n      }\n    }\n  }\n\n  /**\n   * Simple hash-based embedding (fast, not semantic)\n   */\n  simpleEmbed(text: string, dim: number = 256): Float32Array {\n    const embedding = new Float32Array(dim);\n\n    // Multi-pass hash for better distribution\n    for (let i = 0; i < text.length; i++) {\n      const code = text.charCodeAt(i);\n      embedding[i % dim] += code / 255;\n      embedding[(i * 7) % dim] += (code * 0.3) / 255;\n      embedding[(i * 13) % dim] += (code * 0.2) / 255;\n    }\n\n    // Normalize\n    let norm = 0;\n    for (let i = 0; i < dim; i++) {\n      norm += embedding[i] * embedding[i];\n    }\n    norm = Math.sqrt(norm) || 1;\n    for (let i = 0; i < dim; i++) {\n      embedding[i] /= norm;\n    }\n\n    return embedding;\n  }\n\n  /**\n   * Compute cosine similarity between two embeddings\n   */\n  cosineSimilarity(a: Float32Array, b: Float32Array): number {\n    if (ruvectorModule?.cosineSimilarity) {\n      return ruvectorModule.cosineSimilarity(a, b);\n    }\n\n    // JS fallback\n    let dot = 0;\n    let normA = 0;\n    let normB = 0;\n    for (let i = 0; i < a.length; i++) {\n      dot += a[i] * b[i];\n      normA += a[i] * a[i];\n      normB += b[i] * b[i];\n    }\n    return dot / (Math.sqrt(normA) * Math.sqrt(normB) || 1);\n  }\n\n  /**\n   * Get statistics\n   */\n  getStats(): EmbeddingStats {\n    const effective = this.effectiveBackend || this.backend;\n    const ruvectorStats = ruvectorModule?.getStats?.() || {};\n\n    return {\n      backend: this.backend,\n      effectiveBackend: effective,\n      dimension: this.dimension,\n      totalEmbeddings: this.totalEmbeddings,\n      totalLatencyMs: this.totalLatencyMs,\n      avgLatencyMs: this.totalEmbeddings > 0 ? this.totalLatencyMs / this.totalEmbeddings : 0,\n      cacheHits: this.cacheHits,\n      modelLoaded: this.modelLoaded,\n      modelName: effective === 'onnx' ? this.modelName : undefined,\n      simdAvailable: ruvectorStats.simdAvailable ?? onnxAvailable,\n      parallelWorkers: ruvectorStats.workerCount ?? undefined,\n    };\n  }\n\n  /**\n   * Clear cache\n   */\n  clearCache(): void {\n    this.cache.clear();\n  }\n\n  /**\n   * Clear corpus\n   */\n  clearCorpus(): void {\n    this.corpus = { texts: [], embeddings: [] };\n  }\n\n  /**\n   * Shutdown (cleanup workers)\n   */\n  async shutdown(): Promise<void> {\n    if (ruvectorModule?.shutdown) {\n      await ruvectorModule.shutdown();\n    }\n  }\n\n  /**\n   * Reset instance (for testing)\n   */\n  static async reset(): Promise<void> {\n    if (EmbeddingService.instance) {\n      await EmbeddingService.instance.shutdown();\n    }\n    EmbeddingService.instance = null;\n    onnxAvailable = null;\n    ruvectorModule = null;\n  }\n}\n\n// Export singleton getter\nexport function getEmbeddingService(): EmbeddingService {\n  return EmbeddingService.getInstance();\n}\n\n// Export convenience functions\nexport async function embed(text: string): Promise<Float32Array> {\n  return getEmbeddingService().embed(text);\n}\n\nexport async function embedBatch(texts: string[]): Promise<Float32Array[]> {\n  return getEmbeddingService().embedBatch(texts);\n}\n\nexport async function textSimilarity(text1: string, text2: string): Promise<number> {\n  return getEmbeddingService().similarity(text1, text2);\n}\n\nexport function simpleEmbed(text: string, dim: number = 256): Float32Array {\n  return getEmbeddingService().simpleEmbed(text, dim);\n}\n\nexport async function similarityMatrix(texts: string[]): Promise<number[][]> {\n  return getEmbeddingService().similarityMatrix(texts);\n}\n\nexport async function semanticSearch(query: string, topK: number = 5): Promise<SearchResult[]> {\n  return getEmbeddingService().semanticSearch(query, topK);\n}\n\nexport async function findDuplicates(texts: string[], threshold: number = 0.9): Promise<DuplicateGroup[]> {\n  return getEmbeddingService().findDuplicates(texts, threshold);\n}\n\nexport async function clusterTexts(texts: string[], k: number = 3): Promise<{ clusters: number[]; centroids: Float32Array[] }> {\n  return getEmbeddingService().clusterTexts(texts, k);\n}\n"]}

package/dist/intelligence/embedding-benchmark.js

@@ -111,12 +111,16 @@ async function benchmarkService(service, name) {
     const avgWarmLatency = warmTotalLatency / TEST_TEXTS.length;
     console.log(`  Total: ${warmTotalLatency.toFixed(2)}ms, Avg: ${avgWarmLatency.toFixed(2)}ms`);
     // Batch embedding
-    console.log(`[${name}] Batch embedding...`);
+    console.log(`[${name}] Batch embedding (10 texts)...`);
     service.clearCache();
     const batchStart = performance.now();
     await service.embedBatch(TEST_TEXTS);
     const batchLatency = performance.now() - batchStart;
-    console.log(`  Batch latency: ${batchLatency.toFixed(2)}ms`);
+    const batchPerText = batchLatency / TEST_TEXTS.length;
+    console.log(`  Batch latency: ${batchLatency.toFixed(2)}ms (${batchPerText.toFixed(2)}ms per text)`);
+    // Compare batch vs sequential
+    const speedup = avgWarmLatency > 0 ? avgWarmLatency / batchPerText : 0;
+    console.log(`  Batch speedup: ${speedup.toFixed(1)}x vs sequential`);
     // Semantic similarity tests
     console.log(`[${name}] Semantic similarity tests...`);
     const similarities = [];