squish-memory 1.1.5 → 1.2.0

package/dist/core/embeddings/embeddings.js

@@ -0,0 +1,546 @@
+import { config } from '../config.js';
+import { getGoogleMultimodalEmbedding, isMultimodalInput } from './google-multimodal.js';
+import { logger } from '../logger.js';
+// Lazy-import transformers to avoid loading unless requested
+let transformersLocal = null;
+async function getTransformersLocal() {
+    if (!transformersLocal) {
+        transformersLocal = import('./transformers-local.js');
+    }
+    return transformersLocal;
+}
+function missingModelError(provider, envVar) {
+    return new Error(`Embedding provider "${provider}" requires ${envVar} to be set`);
+}
+function requireModel(provider, envVar, model) {
+    if (!model.trim()) {
+        throw missingModelError(provider, envVar);
+    }
+    return model;
+}
+// Retry utility with exponential backoff
+async function withRetry(fn, maxRetries = config.embeddingsMaxRetries, baseDelayMs = config.embeddingsRetryDelayMs) {
+    let lastError;
+    for (let attempt = 0; attempt < maxRetries; attempt++) {
+        try {
+            return await fn();
+        }
+        catch (error) {
+            lastError = error;
+            // Only retry on network errors (5xx, ECONNRESET, ETIMEDOUT, etc.)
+            if (error instanceof Error && shouldRetryError(error)) {
+                const delay = baseDelayMs * Math.pow(2, attempt) + Math.random() * baseDelayMs;
+                logger.debug(`Embedding request failed (attempt ${attempt + 1}/${maxRetries}), retrying in ${delay.toFixed(0)}ms`, { error: error });
+                await new Promise(resolve => setTimeout(resolve, delay));
+                continue;
+            }
+            // Don't retry on 4xx errors or non-retryable errors
+            break;
+        }
+    }
+    throw lastError;
+}
+function shouldRetryError(error) {
+    const message = error.message.toLowerCase();
+    // Network errors that are typically transient
+    const retryablePatterns = [
+        'econnreset',
+        'etimedout',
+        'econnrefused',
+        'esocket',
+        'network error',
+        'fetch failed',
+        'timeout',
+        'request timed out',
+        'service unavailable',
+        'too many requests',
+        'rate limit',
+        'internal server error',
+        'bad gateway',
+        'gateway timeout',
+    ];
+    return retryablePatterns.some(pattern => message.includes(pattern));
+}
+// Timeout wrapper using AbortController
+async function withTimeout(promise, timeoutMs) {
+    const controller = new AbortController();
+    const timeoutId = setTimeout(() => controller.abort(), timeoutMs);
+    try {
+        return await promise;
+    }
+    finally {
+        clearTimeout(timeoutId);
+    }
+}
+// Fetch wrapper that combines retry and timeout
+async function fetchWithRetryAndTimeout(url, options, timeoutMs = config.embeddingsTimeoutMs) {
+    return withRetry(async () => {
+        return withTimeout(fetch(url, options), timeoutMs);
+    });
+}
+// Simple in-memory cache for embeddings (LRU with 1000 entries)
+const embeddingCache = new Map();
+const MAX_CACHE_SIZE = 1000;
+function getCacheKey(input, provider) {
+    // Simple hash of input + provider
+    let hash = 0;
+    const str = input + provider;
+    for (let i = 0; i < str.length; i++) {
+        const char = str.charCodeAt(i);
+        hash = ((hash << 5) - hash) + char;
+        hash = hash & hash; // Convert to 32bit integer
+    }
+    return hash.toString();
+}
+function getCachedEmbedding(key) {
+    return embeddingCache.get(key);
+}
+function setCachedEmbedding(key, embedding) {
+    if (embeddingCache.size >= MAX_CACHE_SIZE) {
+        // Remove oldest entry (first one)
+        const firstKey = embeddingCache.keys().next().value;
+        if (firstKey) {
+            embeddingCache.delete(firstKey);
+        }
+    }
+    embeddingCache.set(key, embedding);
+}
+export async function getEmbedding(input) {
+    if (!input || (typeof input !== 'string' && !isMultimodalInput(input))) {
+        return null;
+    }
+    const provider = config.embeddingsProvider;
+    const cacheKey = typeof input === 'string'
+        ? getCacheKey(input, provider)
+        : getCacheKey(JSON.stringify(input), provider);
+    // Check cache first
+    const cached = getCachedEmbedding(cacheKey);
+    if (cached) {
+        return cached;
+    }
+    let result = null;
+    // Handle multimodal input
+    if (isMultimodalInput(input) && provider === 'google') {
+        requireModel('google', 'SQUISH_GOOGLE_EMBEDDING_MODEL', config.googleEmbeddingModel);
+        const multimodalResult = await getGoogleMultimodalEmbedding(input);
+        if (multimodalResult) {
+            result = multimodalResult.embedding;
+        }
+    }
+    // Handle text-only input
+    if (!result && typeof input === 'string') {
+        const textInput = input;
+        if (provider === 'none') {
+            result = null;
+        }
+        else if (provider === 'google') {
+            requireModel('google', 'SQUISH_GOOGLE_EMBEDDING_MODEL', config.googleEmbeddingModel);
+            const multimodalResult = await getGoogleMultimodalEmbedding({ text: textInput });
+            result = multimodalResult?.embedding || null;
+        }
+        else if (provider === 'openai') {
+            requireModel('openai', 'SQUISH_OPENAI_EMBEDDING_MODEL', config.openAiEmbeddingModel);
+            result = await getOpenAiEmbedding(textInput);
+        }
+        else if (provider === 'ollama') {
+            requireModel('ollama', 'SQUISH_OLLAMA_EMBEDDING_MODEL', config.ollamaEmbeddingModel);
+            result = await getOllamaEmbedding(textInput);
+        }
+        else if (provider === 'lmstudio') {
+            requireModel('lmstudio', 'SQUISH_LM_STUDIO_EMBEDDING_MODEL', config.lmStudioEmbeddingModel);
+            result = await getLmStudioEmbedding(textInput);
+        }
+        else if (provider === 'transformers') {
+            requireModel('transformers', 'SQUISH_LOCAL_MODEL', config.transformersLocalModel);
+            try {
+                const mod = await getTransformersLocal();
+                result = await mod.getEmbedding(textInput);
+            }
+            catch (error) {
+                logger.debug(`Transformers not available, falling back to TF-IDF: ${error}`);
+            }
+            // If transformers failed, use TF-IDF
+            if (!result) {
+                result = getLocalEmbedding(textInput);
+            }
+        }
+        else if (provider === 'local') {
+            result = getLocalEmbedding(textInput);
+        }
+        else {
+            // Auto mode: cloud -> transformers -> TF-IDF (smart fallback)
+            // Step 1: Try cloud providers
+            if (config.openAiApiKey && config.openAiEmbeddingModel) {
+                result = await getOpenAiEmbedding(textInput);
+            }
+            // Step 2: Try Ollama
+            if (!result && config.ollamaUrl && config.ollamaEmbeddingModel) {
+                result = await getOllamaEmbedding(textInput);
+            }
+            // Step 3: Try LM Studio
+            if (!result && config.lmStudioUrl && config.lmStudioEmbeddingModel) {
+                result = await getLmStudioEmbedding(textInput);
+            }
+            // Step 4: Try Transformers.js local
+            if (!result && config.transformersLocalModel) {
+                try {
+                    const mod = await getTransformersLocal();
+                    result = await mod.getEmbedding(textInput);
+                }
+                catch {
+                    // Transformers not available, continue to fallback
+                }
+            }
+            // Step 5: Fall back to TF-IDF (always works)
+            if (!result) {
+                result = getLocalEmbedding(textInput);
+            }
+        }
+    }
+    // Cache the result if valid
+    if (result) {
+        setCachedEmbedding(cacheKey, result);
+    }
+    return result;
+}
+/**
+ * Get embeddings for multiple inputs in parallel batches
+ * Processes inputs in batches to respect rate limits while parallelizing
+ */
+export async function getBatchEmbeddings(inputs, batchSize = 20) {
+    if (inputs.length === 0)
+        return [];
+    const results = new Array(inputs.length).fill(null);
+    const provider = config.embeddingsProvider;
+    // Check cache for all inputs first
+    const uncachedIndices = [];
+    const uncachedInputs = [];
+    for (let i = 0; i < inputs.length; i++) {
+        const cacheKey = getCacheKey(inputs[i], provider);
+        const cached = getCachedEmbedding(cacheKey);
+        if (cached) {
+            results[i] = cached;
+        }
+        else {
+            uncachedIndices.push(i);
+            uncachedInputs.push(inputs[i]);
+        }
+    }
+    // Process only uncached inputs in batches
+    for (let i = 0; i < uncachedInputs.length; i += batchSize) {
+        const batchEnd = Math.min(i + batchSize, uncachedInputs.length);
+        const batch = uncachedInputs.slice(i, batchEnd);
+        const indices = uncachedIndices.slice(i, batchEnd);
+        // Parallelize embeddings within batch using Promise.all
+        const batchResults = await Promise.all(batch.map((input) => getEmbedding(input)));
+        // Store results in correct positions and cache
+        for (let j = 0; j < batchResults.length; j++) {
+            results[indices[j]] = batchResults[j];
+        }
+    }
+    return results;
+}
+/**
+ * Clear the embedding cache
+ */
+export function clearEmbeddingCache() {
+    embeddingCache.clear();
+}
+/**
+ * Get embedding cache statistics
+ */
+export function getEmbeddingCacheStats() {
+    return {
+        size: embeddingCache.size,
+        maxSize: MAX_CACHE_SIZE,
+    };
+}
+/**
+ * Local TF-IDF embedding using character n-grams and word hashing
+ * Creates a 768-dimensional vector for fast offline similarity.
+ * Fast, no API calls, works offline
+ */
+function getLocalEmbedding(input) {
+    if (!input || typeof input !== 'string') {
+        return null;
+    }
+    // Normalize text
+    const text = input.toLowerCase()
+        .replace(/[^a-z0-9\s]/g, ' ')
+        .replace(/\s+/g, ' ')
+        .trim();
+    if (text.length === 0) {
+        return null;
+    }
+    // Embedding dimensions
+    const dimensions = 768;
+    const vector = new Array(dimensions).fill(0);
+    // Character n-grams (3-5 grams for semantic similarity)
+    const ngrams = [3, 4, 5];
+    for (const n of ngrams) {
+        for (let i = 0; i <= text.length - n; i++) {
+            const gram = text.substring(i, i + n);
+            const hash = djb2Hash(gram);
+            const idx = Math.abs(hash) % dimensions;
+            vector[idx] += 1;
+        }
+    }
+    // Word-level hashing for semantic capture
+    const words = text.split(/\s+/).filter(w => w.length > 2);
+    for (const word of words) {
+        const hash = djb2Hash(word);
+        const idx = Math.abs(hash) % dimensions;
+        vector[idx] += 2; // Words weighted higher than n-grams
+        // Bigrams
+        if (words.length > 1) {
+            const idx2 = words.indexOf(word);
+            if (idx2 < words.length - 1) {
+                const bigram = `${word}_${words[idx2 + 1]}`;
+                const bigramHash = djb2Hash(bigram);
+                const bigramIdx = Math.abs(bigramHash) % dimensions;
+                vector[bigramIdx] += 3; // Bigrams weighted highest
+            }
+        }
+    }
+    // Apply TF-IDF-like scaling: square root to dampen high frequencies
+    for (let i = 0; i < dimensions; i++) {
+        if (vector[i] > 0) {
+            vector[i] = Math.sqrt(vector[i]);
+        }
+    }
+    // L2 normalize
+    const norm = Math.sqrt(vector.reduce((sum, val) => sum + val * val, 0));
+    if (norm > 0) {
+        for (let i = 0; i < dimensions; i++) {
+            vector[i] /= norm;
+        }
+    }
+    return vector;
+}
+/**
+ * DJB2 hash function - fast, good distribution
+ */
+function djb2Hash(str) {
+    let hash = 5381;
+    for (let i = 0; i < str.length; i++) {
+        hash = ((hash << 5) + hash) + str.charCodeAt(i);
+    }
+    return hash;
+}
+async function getOpenAiEmbedding(input) {
+    if (!config.openAiApiKey)
+        return null;
+    if (!config.openAiEmbeddingModel)
+        return null;
+    try {
+        const response = await fetchWithRetryAndTimeout(config.openAiApiUrl, {
+            method: 'POST',
+            headers: {
+                'Content-Type': 'application/json',
+                Authorization: `Bearer ${config.openAiApiKey}`,
+            },
+            body: JSON.stringify({
+                model: config.openAiEmbeddingModel,
+                input,
+            }),
+        }, config.openAiTimeoutMs);
+        if (!response.ok) {
+            const message = await response.text();
+            logger.warn(`OpenAI embeddings failed: ${response.status} ${message}`);
+            return null; // Return null to allow fallback
+        }
+        const payload = await response.json();
+        const embedding = payload.data?.[0]?.embedding;
+        return embedding ?? null;
+    }
+    catch (error) {
+        logger.warn('OpenAI embeddings error:', { error: error });
+        return null; // Return null to allow fallback
+    }
+}
+async function getOllamaEmbedding(input) {
+    if (!config.ollamaEmbeddingModel)
+        return null;
+    try {
+        const response = await fetchWithRetryAndTimeout(`${config.ollamaUrl}/api/embeddings`, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({
+                model: config.ollamaEmbeddingModel,
+                prompt: input,
+            }),
+        }, config.ollamaTimeoutMs);
+        if (!response.ok) {
+            const message = await response.text();
+            logger.warn(`Ollama embeddings failed: ${response.status} ${message}`);
+            return null; // Return null to allow fallback
+        }
+        const payload = await response.json();
+        return payload.embedding ?? null;
+    }
+    catch (error) {
+        logger.warn('Ollama embeddings error:', { error: error });
+        return null; // Return null to allow fallback
+    }
+}
+// LM Studio uses OpenAI-compatible API
+async function getLmStudioEmbedding(input) {
+    if (!config.lmStudioEmbeddingModel)
+        return null;
+    try {
+        const response = await fetchWithRetryAndTimeout(`${config.lmStudioUrl}/v1/embeddings`, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({
+                model: config.lmStudioEmbeddingModel,
+                input: input,
+            }),
+        }, config.ollamaTimeoutMs); // Reuse Ollama timeout
+        if (!response.ok) {
+            const message = await response.text();
+            logger.warn(`LM Studio embeddings failed: ${response.status} ${message}`);
+            return null; // Return null to allow fallback
+        }
+        const payload = await response.json();
+        return payload.data?.[0]?.embedding ?? null;
+    }
+    catch (error) {
+        logger.warn('LM Studio embeddings error:', { error: error });
+        return null; // Return null to allow fallback
+    }
+}
+/**
+ * Check health of all configured embedding providers
+ * Returns availability and latency for each provider
+ */
+export async function checkEmbeddingProviderHealth() {
+    const results = new Map();
+    const providers = ['local', 'openai', 'ollama', 'lmstudio', 'transformers', 'google', 'none', 'auto'];
+    // Test local provider (always available)
+    results.set('local', { available: true, latencyMs: 0 });
+    // Test OpenAI if configured
+    if (config.openAiApiKey && config.openAiEmbeddingModel) {
+        const start = Date.now();
+        try {
+            const testInput = 'health check';
+            const embedding = await withRetry(() => withTimeout(getOpenAiEmbedding(testInput), config.openAiTimeoutMs), config.embeddingsMaxRetries, config.embeddingsRetryDelayMs);
+            const latency = Date.now() - start;
+            results.set('openai', {
+                available: embedding !== null && embedding.length > 0,
+                latencyMs: latency
+            });
+        }
+        catch (error) {
+            results.set('openai', {
+                available: false,
+                error: error.message
+            });
+        }
+    }
+    else {
+        results.set('openai', { available: false, error: 'Not configured' });
+    }
+    // Test Ollama if configured
+    if (config.ollamaUrl && config.ollamaEmbeddingModel) {
+        const start = Date.now();
+        try {
+            const testInput = 'health check';
+            const embedding = await withRetry(() => withTimeout(getOllamaEmbedding(testInput), config.ollamaTimeoutMs), config.embeddingsMaxRetries, config.embeddingsRetryDelayMs);
+            const latency = Date.now() - start;
+            results.set('ollama', {
+                available: embedding !== null && embedding.length > 0,
+                latencyMs: latency
+            });
+        }
+        catch (error) {
+            results.set('ollama', {
+                available: false,
+                error: error.message
+            });
+        }
+    }
+    else {
+        results.set('ollama', { available: false, error: 'Not configured' });
+    }
+    // Test LM Studio if configured
+    if (config.lmStudioUrl && config.lmStudioEmbeddingModel) {
+        const start = Date.now();
+        try {
+            const testInput = 'health check';
+            const embedding = await withRetry(() => withTimeout(getLmStudioEmbedding(testInput), config.ollamaTimeoutMs), config.embeddingsMaxRetries, config.embeddingsRetryDelayMs);
+            const latency = Date.now() - start;
+            results.set('lmstudio', {
+                available: embedding !== null && embedding.length > 0,
+                latencyMs: latency
+            });
+        }
+        catch (error) {
+            results.set('lmstudio', {
+                available: false,
+                error: error.message
+            });
+        }
+    }
+    else {
+        results.set('lmstudio', { available: false, error: 'Not configured' });
+    }
+    // Test Transformers.js local if requested
+    const transformersHealth = async () => {
+        try {
+            const mod = await getTransformersLocal();
+            const health = await mod.checkHealth();
+            return health;
+        }
+        catch (error) {
+            return { available: false, error: error.message };
+        }
+    };
+    // Try to test transformers (library must be installed)
+    if (!config.transformersLocalModel) {
+        results.set('transformers', {
+            available: false,
+            error: 'Not configured',
+        });
+    }
+    else
+        try {
+            const start = Date.now();
+            const mod = await getTransformersLocal();
+            const health = await mod.checkHealth();
+            const latency = Date.now() - start;
+            results.set('transformers', {
+                available: health.available,
+                latencyMs: latency,
+                error: health.error,
+            });
+        }
+        catch (error) {
+            results.set('transformers', {
+                available: false,
+                error: error.message,
+            });
+        }
+    // Test Google if configured
+    if ((config.googleCloudApiKey || config.googleCloudProject) && config.googleEmbeddingModel) {
+        const start = Date.now();
+        try {
+            const result = await withRetry(() => withTimeout(getGoogleMultimodalEmbedding({ text: 'health check' }), config.googleTimeoutMs), config.embeddingsMaxRetries, config.embeddingsRetryDelayMs);
+            const latency = Date.now() - start;
+            results.set('google', {
+                available: result !== null && result.embedding.length > 0,
+                latencyMs: latency
+            });
+        }
+        catch (error) {
+            results.set('google', {
+                available: false,
+                error: error.message
+            });
+        }
+    }
+    else {
+        results.set('google', { available: false, error: 'Not configured' });
+    }
+    return results;
+}
+//# sourceMappingURL=embeddings.js.map

package/dist/core/embeddings/google-multimodal.js

@@ -5,10 +5,14 @@ export async function getGoogleMultimodalEmbedding(input) {
         logger.debug('Google Cloud credentials not configured');
         return null;
     }
+    if (!config.googleEmbeddingModel) {
+        logger.debug('Google embedding model not configured');
+        return null;
+    }
     try {
         const endpoint = config.googleCloudLocation === 'global'
-            ? `https://aiplatform.googleapis.com/v1/projects/${config.googleCloudProject}/locations/${config.googleCloudLocation}/publishers/google/models/multimodalembedding@001:predict`
-            : `https://${config.googleCloudLocation}-aiplatform.googleapis.com/v1/projects/${config.googleCloudProject}/locations/${config.googleCloudLocation}/publishers/google/models/multimodalembedding@001:predict`;
+            ? `https://aiplatform.googleapis.com/v1/projects/${config.googleCloudProject}/locations/${config.googleCloudLocation}/publishers/google/models/${config.googleEmbeddingModel}:predict`
+            : `https://${config.googleCloudLocation}-aiplatform.googleapis.com/v1/projects/${config.googleCloudProject}/locations/${config.googleCloudLocation}/publishers/google/models/${config.googleEmbeddingModel}:predict`;
         const instances = [];
         if (input.text) {
             instances.push({ text: input.text });

package/dist/core/{local-embeddings.d.ts → embeddings/local-embeddings.d.ts}

@@ -2,7 +2,7 @@
  * Utility functions for local embeddings
  * Note: Actual embedding generation is in core/embeddings.ts
  */
-import { cosineSimilarity as vectorCosineSimilarity } from './utils/vector-operations.js';
+import { cosineSimilarity as vectorCosineSimilarity } from '../utils/vector-operations.js';
 /**
  * @deprecated Use cosineSimilarity from core/utils/vector-operations.ts directly.
  *   This re-export is for backward compatibility and will be removed in v1.2.0.

package/dist/core/embeddings/local-embeddings.js

@@ -0,0 +1,11 @@
+/**
+ * Utility functions for local embeddings
+ * Note: Actual embedding generation is in core/embeddings.ts
+ */
+import { cosineSimilarity as vectorCosineSimilarity } from '../utils/vector-operations.js';
+/**
+ * @deprecated Use cosineSimilarity from core/utils/vector-operations.ts directly.
+ *   This re-export is for backward compatibility and will be removed in v1.2.0.
+ */
+export const cosineSimilarity = vectorCosineSimilarity;
+//# sourceMappingURL=local-embeddings.js.map

package/dist/core/embeddings/qmd-client.js

@@ -18,7 +18,7 @@
 import { Client } from '@modelcontextprotocol/sdk/client/index.js';
 import { StdioClientTransport } from '@modelcontextprotocol/sdk/client/stdio.js';
 import { spawn } from 'child_process';
-import { logger } from '../logger.js';
+import { logger } from '../logger';
 /**
  * QMD MCP Client class
  *

package/dist/core/embeddings/transformers-local.d.ts

@@ -0,0 +1,64 @@
+/**
+ * Transformers.js Local Embedding Provider
+ *
+ * Uses ONNX-based transformer models for high-quality local embeddings.
+ * Supports Hugging Face ONNX embedding models.
+ *
+ * Usage:
+ *   Set SQUISH_EMBEDDINGS_PROVIDER=transformers
+ *   Required: SQUISH_LOCAL_MODEL=<huggingface-onnx-model>
+ *
+ * Download models automatically on first use. Models cached in HuggingFace cache directory.
+ */
+export interface TransformersLocalConfig {
+    model: string;
+    device: 'cpu' | 'webgpu';
+    dtype: 'q8' | 'q4' | 'f16' | 'f32';
+}
+/**
+ * Check if pipeline is loaded
+ */
+export declare function isReady(): boolean;
+/**
+ * Get embedding dimension for current model
+ */
+export declare function getEmbeddingDimension(): number;
+/**
+ * Generate embedding for a single text input
+ * Uses mean pooling + L2 normalization
+ */
+export declare function getEmbedding(text: string): Promise<number[] | null>;
+/**
+ * Generate embeddings for multiple texts in batch
+ * Processes efficiently with batching
+ */
+export declare function getBatchEmbeddings(texts: string[], batchSize?: number): Promise<Array<number[] | null>>;
+/**
+ * Check health of the transformers provider
+ */
+export declare function checkHealth(): Promise<{
+    available: boolean;
+    latencyMs?: number;
+    error?: string;
+    model?: string;
+    dimension?: number;
+}>;
+/**
+ * Unload the pipeline (for testing or memory management)
+ */
+export declare function unload(): Promise<void>;
+/**
+ * Warm up the model with a test input
+ */
+export declare function warmup(): Promise<boolean>;
+declare const _default: {
+    isReady: typeof isReady;
+    getEmbeddingDimension: typeof getEmbeddingDimension;
+    getEmbedding: typeof getEmbedding;
+    getBatchEmbeddings: typeof getBatchEmbeddings;
+    checkHealth: typeof checkHealth;
+    unload: typeof unload;
+    warmup: typeof warmup;
+};
+export default _default;
+//# sourceMappingURL=transformers-local.d.ts.map