recallnest 1.4.1

package/src/embedder.ts

@@ -0,0 +1,557 @@
+/**
+ * Embedding Abstraction Layer
+ * OpenAI-compatible API for various embedding providers.
+ * Supports automatic chunking for documents exceeding embedding context limits.
+ *
+ * Note: Some providers (e.g. Jina) support extra parameters like `task` and
+ * `normalized` on the embeddings endpoint. The OpenAI SDK types do not include
+ * these fields, so we pass them via a narrow `any` cast.
+ */
+
+import OpenAI from "openai";
+import { createHash } from "node:crypto";
+import { smartChunk } from "./chunker.js";
+import { logInfo, logWarn } from "./stderr-log.js";
+
+const EMBEDDING_RETRY_DELAYS_MS = [200, 500];
+const EMBEDDING_MAX_CONCURRENCY = 4;
+
+// ============================================================================
+// Embedding Cache (LRU with TTL)
+// ============================================================================
+
+interface CacheEntry {
+  vector: number[];
+  createdAt: number;
+}
+
+class EmbeddingCache {
+  private cache = new Map<string, CacheEntry>();
+  private readonly maxSize: number;
+  private readonly ttlMs: number;
+  public hits = 0;
+  public misses = 0;
+
+  constructor(maxSize = 256, ttlMinutes = 30) {
+    this.maxSize = maxSize;
+    this.ttlMs = ttlMinutes * 60_000;
+  }
+
+  private key(text: string, task?: string): string {
+    const hash = createHash("sha256").update(`${task || ""}:${text}`).digest("hex").slice(0, 24);
+    return hash;
+  }
+
+  get(text: string, task?: string): number[] | undefined {
+    const k = this.key(text, task);
+    const entry = this.cache.get(k);
+    if (!entry) {
+      this.misses++;
+      return undefined;
+    }
+    if (Date.now() - entry.createdAt > this.ttlMs) {
+      this.cache.delete(k);
+      this.misses++;
+      return undefined;
+    }
+    // Move to end (most recently used)
+    this.cache.delete(k);
+    this.cache.set(k, entry);
+    this.hits++;
+    return entry.vector;
+  }
+
+  set(text: string, task: string | undefined, vector: number[]): void {
+    const k = this.key(text, task);
+    // Evict oldest if full
+    if (this.cache.size >= this.maxSize) {
+      const firstKey = this.cache.keys().next().value;
+      if (firstKey !== undefined) this.cache.delete(firstKey);
+    }
+    this.cache.set(k, { vector, createdAt: Date.now() });
+  }
+
+  get size(): number { return this.cache.size; }
+  get stats(): { size: number; hits: number; misses: number; hitRate: string } {
+    const total = this.hits + this.misses;
+    return {
+      size: this.cache.size,
+      hits: this.hits,
+      misses: this.misses,
+      hitRate: total > 0 ? `${((this.hits / total) * 100).toFixed(1)}%` : "N/A",
+    };
+  }
+}
+
+// ============================================================================
+// Types & Configuration
+// ============================================================================
+
+export interface EmbeddingConfig {
+  provider: "openai-compatible";
+  /** API key. Optional for local providers (Ollama, Docker Model Runner). */
+  apiKey?: string;
+  model: string;
+  baseURL?: string;
+  dimensions?: number;
+
+  /** Optional task type for query embeddings (e.g. "retrieval.query") */
+  taskQuery?: string;
+  /** Optional task type for passage/document embeddings (e.g. "retrieval.passage") */
+  taskPassage?: string;
+  /** Optional flag to request normalized embeddings (provider-dependent, e.g. Jina v5) */
+  normalized?: boolean;
+  /** Enable automatic chunking for documents exceeding context limits (default: true) */
+  chunking?: boolean;
+}
+
+// Known embedding model dimensions
+const EMBEDDING_DIMENSIONS: Record<string, number> = {
+  "text-embedding-3-small": 1536,
+  "text-embedding-3-large": 3072,
+  "text-embedding-004": 768,
+  "gemini-embedding-001": 3072,
+  "nomic-embed-text": 768,
+  "mxbai-embed-large": 1024,
+  "BAAI/bge-m3": 1024,
+  "all-MiniLM-L6-v2": 384,
+  "all-mpnet-base-v2": 512,
+
+  // Jina v5
+  "jina-embeddings-v5-text-small": 1024,
+  "jina-embeddings-v5-text-nano": 768,
+
+  // Qwen3 Embedding (Docker Model Runner / local)
+  "ai/qwen3-embedding": 1024,
+  "ai/qwen3-embedding:0.6B-F16": 1024,
+  "ai/qwen3-embedding:4B": 1024,
+  "ai/qwen3-embedding:4B-Q4_K_M": 1024,
+  "ai/qwen3-embedding:8B-Q4_K_M": 1024,
+};
+
+// ============================================================================
+// Utility Functions
+// ============================================================================
+
+function resolveEnvVars(value: string): string {
+  return value.replace(/\$\{([^}]+)\}/g, (_, envVar) => {
+    const envValue = process.env[envVar];
+    if (!envValue) {
+      throw new Error(`Environment variable ${envVar} is not set`);
+    }
+    return envValue;
+  });
+}
+
+function sleep(ms: number): Promise<void> {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+function isTransientEmbeddingError(error: unknown): boolean {
+  if (error instanceof OpenAI.APIConnectionError || error instanceof OpenAI.APIConnectionTimeoutError) {
+    return true;
+  }
+
+  const message = error instanceof Error ? error.message : String(error);
+  return /ECONNRESET|socket connection was closed unexpectedly|connection error|timed out|network/i.test(message);
+}
+
+export function getVectorDimensions(model: string, overrideDims?: number): number {
+  if (overrideDims && overrideDims > 0) {
+    return overrideDims;
+  }
+
+  const dims = EMBEDDING_DIMENSIONS[model];
+  if (!dims) {
+    throw new Error(
+      `Unsupported embedding model: ${model}. Either add it to EMBEDDING_DIMENSIONS or set embedding.dimensions in config.`
+    );
+  }
+
+  return dims;
+}
+
+// ============================================================================
+// Embedder Class
+// ============================================================================
+
+export class Embedder {
+  private client: OpenAI;
+  public readonly dimensions: number;
+  private readonly _cache: EmbeddingCache;
+  private readonly _maxConcurrentRequests = EMBEDDING_MAX_CONCURRENCY;
+  private _activeRequests = 0;
+  private readonly _pendingRequests: Array<() => void> = [];
+
+  private readonly _model: string;
+  private readonly _taskQuery?: string;
+  private readonly _taskPassage?: string;
+  private readonly _normalized?: boolean;
+
+  /** Optional requested dimensions to pass through to the embedding provider (OpenAI-compatible). */
+  private readonly _requestDimensions?: number;
+  /** Enable automatic chunking for long documents (default: true) */
+  private readonly _autoChunk: boolean;
+
+  constructor(config: EmbeddingConfig & { chunking?: boolean }) {
+    // Resolve environment variables in API key (fallback for local providers)
+    const resolvedApiKey = config.apiKey ? resolveEnvVars(config.apiKey) : "no-key-required";
+
+    this._model = config.model;
+    this._taskQuery = config.taskQuery;
+    this._taskPassage = config.taskPassage;
+    this._normalized = config.normalized;
+    this._requestDimensions = config.dimensions;
+    // Enable auto-chunking by default for better handling of long documents
+    this._autoChunk = config.chunking !== false;
+
+    this.client = new OpenAI({
+      apiKey: resolvedApiKey,
+      ...(config.baseURL ? { baseURL: config.baseURL } : {}),
+    });
+
+    this.dimensions = getVectorDimensions(config.model, config.dimensions);
+    this._cache = new EmbeddingCache(256, 30); // 256 entries, 30 min TTL
+  }
+
+  // --------------------------------------------------------------------------
+  // Backward-compatible API
+  // --------------------------------------------------------------------------
+
+  /**
+   * Backward-compatible embedding API.
+   *
+   * Historically the plugin used a single `embed()` method for both query and
+   * passage embeddings. With task-aware providers we treat this as passage.
+   */
+  async embed(text: string): Promise<number[]> {
+    return this.embedPassage(text);
+  }
+
+  /** Backward-compatible batch embedding API (treated as passage). */
+  async embedBatch(texts: string[]): Promise<number[][]> {
+    return this.embedBatchPassage(texts);
+  }
+
+  // --------------------------------------------------------------------------
+  // Task-aware API
+  // --------------------------------------------------------------------------
+
+  async embedQuery(text: string): Promise<number[]> {
+    return this.embedSingle(text, this._taskQuery);
+  }
+
+  async embedPassage(text: string): Promise<number[]> {
+    return this.embedSingle(text, this._taskPassage);
+  }
+
+  async embedBatchQuery(texts: string[]): Promise<number[][]> {
+    return this.embedMany(texts, this._taskQuery);
+  }
+
+  async embedBatchPassage(texts: string[]): Promise<number[][]> {
+    return this.embedMany(texts, this._taskPassage);
+  }
+
+  // --------------------------------------------------------------------------
+  // Internals
+  // --------------------------------------------------------------------------
+
+  private validateEmbedding(embedding: number[]): void {
+    if (!Array.isArray(embedding)) {
+      throw new Error(`Embedding is not an array (got ${typeof embedding})`);
+    }
+    if (embedding.length !== this.dimensions) {
+      throw new Error(
+        `Embedding dimension mismatch: expected ${this.dimensions}, got ${embedding.length}`
+      );
+    }
+  }
+
+  private buildPayload(input: string | string[], task?: string): any {
+    const payload: any = {
+      model: this.model,
+      input,
+      // Force float output to avoid SDK default base64 decoding path.
+      encoding_format: "float",
+    };
+
+    if (task) payload.task = task;
+    if (this._normalized !== undefined) payload.normalized = this._normalized;
+
+    // Some OpenAI-compatible providers support requesting a specific vector size.
+    // We only pass it through when explicitly configured to avoid breaking providers
+    // that reject unknown fields.
+    if (this._requestDimensions && this._requestDimensions > 0) {
+      payload.dimensions = this._requestDimensions;
+    }
+
+    return payload;
+  }
+
+  private async withEmbeddingSlot<T>(fn: () => Promise<T>): Promise<T> {
+    if (this._activeRequests >= this._maxConcurrentRequests) {
+      await new Promise<void>((resolve) => {
+        this._pendingRequests.push(resolve);
+      });
+    }
+
+    this._activeRequests += 1;
+    try {
+      return await fn();
+    } finally {
+      this._activeRequests = Math.max(0, this._activeRequests - 1);
+      const next = this._pendingRequests.shift();
+      next?.();
+    }
+  }
+
+  private async createEmbedding(input: string | string[], task?: string) {
+    let attempt = 0;
+
+    while (true) {
+      try {
+        return await this.withEmbeddingSlot(() =>
+          this.client.embeddings.create(this.buildPayload(input, task) as any)
+        );
+      } catch (error) {
+        if (!isTransientEmbeddingError(error) || attempt >= EMBEDDING_RETRY_DELAYS_MS.length) {
+          throw error;
+        }
+
+        const delayMs = EMBEDDING_RETRY_DELAYS_MS[attempt];
+        const errorMsg = error instanceof Error ? error.message : String(error);
+        logWarn(`Transient embedding error, retrying in ${delayMs}ms (attempt ${attempt + 1}/${EMBEDDING_RETRY_DELAYS_MS.length})`, errorMsg);
+        attempt += 1;
+        await sleep(delayMs);
+      }
+    }
+  }
+
+  private async embedSingle(text: string, task?: string): Promise<number[]> {
+    if (!text || text.trim().length === 0) {
+      throw new Error("Cannot embed empty text");
+    }
+
+    // Check cache first
+    const cached = this._cache.get(text, task);
+    if (cached) return cached;
+
+    try {
+      const response = await this.createEmbedding(text, task);
+      const embedding = response.data[0]?.embedding as number[] | undefined;
+      if (!embedding) {
+        throw new Error("No embedding returned from provider");
+      }
+
+      this.validateEmbedding(embedding);
+      this._cache.set(text, task, embedding);
+      return embedding;
+    } catch (error) {
+      // Check if this is a context length exceeded error and try chunking
+      const errorMsg = error instanceof Error ? error.message : String(error);
+      const isContextError = /context|too long|exceed|length/i.test(errorMsg);
+
+      if (isContextError && this._autoChunk) {
+        try {
+          logInfo(`Document exceeded context limit (${errorMsg}), attempting chunking...`);
+          const chunkResult = smartChunk(text, this._model);
+          
+          if (chunkResult.chunks.length === 0) {
+            throw new Error(`Failed to chunk document: ${errorMsg}`);
+          }
+
+          // Embed all chunks in parallel
+          logInfo(`Split document into ${chunkResult.chunkCount} chunks for embedding`);
+          const chunkEmbeddings = await Promise.all(
+            chunkResult.chunks.map(async (chunk, idx) => {
+              try {
+                const embedding = await this.embedSingle(chunk, task);
+                return { embedding };
+              } catch (chunkError) {
+                logWarn(`Failed to embed chunk ${idx}:`, chunkError);
+                throw chunkError;
+              }
+            })
+          );
+
+          // Compute average embedding across chunks
+          const avgEmbedding = chunkEmbeddings.reduce(
+            (sum, { embedding }) => {
+              for (let i = 0; i < embedding.length; i++) {
+                sum[i] += embedding[i];
+              }
+              return sum;
+            },
+            new Array(this.dimensions).fill(0)
+          );
+
+          const finalEmbedding = avgEmbedding.map(v => v / chunkEmbeddings.length);
+          
+          // Cache the result for the original text (using its hash)
+          this._cache.set(text, task, finalEmbedding);
+          logInfo(`Successfully embedded long document as ${chunkEmbeddings.length} averaged chunks`);
+          
+          return finalEmbedding;
+        } catch (chunkError) {
+          // If chunking fails, throw the original error
+          logWarn(`Chunking failed, using original error:`, chunkError);
+          throw new Error(`Failed to generate embedding: ${errorMsg}`, { cause: error });
+        }
+      }
+
+      if (error instanceof Error) {
+        throw new Error(`Failed to generate embedding: ${error.message}`, { cause: error });
+      }
+      throw new Error(`Failed to generate embedding: ${String(error)}`);
+    }
+  }
+
+  private async embedMany(texts: string[], task?: string): Promise<number[][]> {
+    if (!texts || texts.length === 0) {
+      return [];
+    }
+
+    // Filter out empty texts and track indices
+    const validTexts: string[] = [];
+    const validIndices: number[] = [];
+
+    texts.forEach((text, index) => {
+      if (text && text.trim().length > 0) {
+        validTexts.push(text);
+        validIndices.push(index);
+      }
+    });
+
+    if (validTexts.length === 0) {
+      return texts.map(() => []);
+    }
+
+    try {
+      const response = await this.createEmbedding(validTexts, task);
+
+      // Create result array with proper length
+      const results: number[][] = new Array(texts.length);
+
+      // Fill in embeddings for valid texts
+      response.data.forEach((item, idx) => {
+        const originalIndex = validIndices[idx];
+        const embedding = item.embedding as number[];
+
+        this.validateEmbedding(embedding);
+        results[originalIndex] = embedding;
+      });
+
+      // Fill empty arrays for invalid texts
+      for (let i = 0; i < texts.length; i++) {
+        if (!results[i]) {
+          results[i] = [];
+        }
+      }
+
+      return results;
+    } catch (error) {
+      // Check if this is a context length exceeded error and try chunking each text
+      const errorMsg = error instanceof Error ? error.message : String(error);
+      const isContextError = /context|too long|exceed|length/i.test(errorMsg);
+
+      if (isContextError && this._autoChunk) {
+        try {
+          logInfo(`Batch embedding failed with context error, attempting chunking...`);
+          
+          const chunkResults = await Promise.all(
+            validTexts.map(async (text, idx) => {
+              const chunkResult = smartChunk(text, this._model);
+              if (chunkResult.chunks.length === 0) {
+                throw new Error("Chunker produced no chunks");
+              }
+
+              // Embed all chunks in parallel, then average.
+              const embeddings = await Promise.all(
+                chunkResult.chunks.map((chunk) => this.embedSingle(chunk, task))
+              );
+
+              const avgEmbedding = embeddings.reduce(
+                (sum, emb) => {
+                  for (let i = 0; i < emb.length; i++) {
+                    sum[i] += emb[i];
+                  }
+                  return sum;
+                },
+                new Array(this.dimensions).fill(0)
+              );
+
+              const finalEmbedding = avgEmbedding.map((v) => v / embeddings.length);
+
+              // Cache the averaged embedding for the original (long) text.
+              this._cache.set(text, task, finalEmbedding);
+
+              return { embedding: finalEmbedding, index: validIndices[idx] };
+            })
+          );
+
+          logInfo(`Successfully chunked and embedded ${chunkResults.length} long documents`);
+
+          // Build results array
+          const results: number[][] = new Array(texts.length);
+          chunkResults.forEach(({ embedding, index }) => {
+            if (embedding.length > 0) {
+              this.validateEmbedding(embedding);
+              results[index] = embedding;
+            } else {
+              results[index] = [];
+            }
+          });
+
+          // Fill empty arrays for invalid texts
+          for (let i = 0; i < texts.length; i++) {
+            if (!results[i]) {
+              results[i] = [];
+            }
+          }
+
+          return results;
+        } catch (chunkError) {
+          throw new Error(`Failed to embed documents after chunking attempt: ${errorMsg}`);
+        }
+      }
+
+      if (error instanceof Error) {
+        throw new Error(`Failed to generate batch embeddings: ${error.message}`, { cause: error });
+      }
+      throw new Error(`Failed to generate batch embeddings: ${String(error)}`);
+    }
+  }
+
+  get model(): string {
+    return this._model;
+  }
+
+  // Test connection and validate configuration
+  async test(): Promise<{ success: boolean; error?: string; dimensions?: number }> {
+    try {
+      const testEmbedding = await this.embedPassage("test");
+      return {
+        success: true,
+        dimensions: testEmbedding.length,
+      };
+    } catch (error) {
+      return {
+        success: false,
+        error: error instanceof Error ? error.message : String(error),
+      };
+    }
+  }
+
+  get cacheStats() {
+    return this._cache.stats;
+  }
+}
+
+// ============================================================================
+// Factory Function
+// ============================================================================
+
+export function createEmbedder(config: EmbeddingConfig): Embedder {
+  return new Embedder(config);
+}

package/src/entity-resolver.ts

@@ -0,0 +1,131 @@
+/**
+ * Entity Resolver — normalize tech entity aliases to canonical names.
+ *
+ * Borrowed from UltraMemory's entity-resolver.ts, adapted for RecallNest:
+ * - Extended with Chinese aliases (RecallNest users are bilingual)
+ * - Hooked into query-expander as a pre-normalization step
+ * - Also usable at capture time to normalize entities before canonical key generation
+ *
+ * Pure mapping logic, zero API calls.
+ */
+
+// ---------------------------------------------------------------------------
+// Built-in aliases (lowercase key → canonical name)
+// ---------------------------------------------------------------------------
+
+export const BUILTIN_ALIASES: Record<string, string> = {
+  // Programming languages
+  "ts": "typescript", "typescript": "typescript",
+  "js": "javascript", "javascript": "javascript",
+  "py": "python", "python": "python",
+  "rb": "ruby", "ruby": "ruby",
+  "rs": "rust", "rust": "rust",
+  "go": "golang", "golang": "golang",
+  "cs": "csharp", "c#": "csharp",
+
+  // Runtimes
+  "node.js": "nodejs", "node": "nodejs", "nodejs": "nodejs",
+  "bun": "bun", "deno": "deno",
+
+  // Frameworks
+  "react.js": "react", "reactjs": "react", "react": "react",
+  "vue.js": "vue", "vuejs": "vue", "vue": "vue",
+  "next.js": "nextjs", "next": "nextjs", "nextjs": "nextjs",
+  "express.js": "express", "express": "express",
+  "angular.js": "angular", "angularjs": "angular", "angular": "angular",
+  "svelte": "svelte", "sveltekit": "sveltekit",
+
+  // Infrastructure
+  "k8s": "kubernetes", "kubernetes": "kubernetes",
+  "pg": "postgresql", "postgres": "postgresql", "postgresql": "postgresql",
+  "mongo": "mongodb", "mongodb": "mongodb",
+  "redis": "redis",
+  "docker": "docker",
+  "docker-compose": "docker compose", "docker compose": "docker compose",
+  "tf": "terraform", "terraform": "terraform",
+  "gcp": "google cloud platform",
+  "aws": "amazon web services",
+
+  // AI/ML
+  "gpt": "openai gpt", "chatgpt": "openai gpt",
+  "claude": "anthropic claude", "anthropic": "anthropic claude",
+  "gemini": "google gemini",
+  "llm": "large language model",
+  "rag": "retrieval augmented generation",
+  "lancedb": "lancedb",
+  "pt": "pytorch", "pytorch": "pytorch",
+  "tf.js": "tensorflow", "tensorflow": "tensorflow",
+  "hf": "hugging face", "huggingface": "hugging face",
+
+  // Project-specific
+  "recallnest": "recallnest", "recall nest": "recallnest", "recall-nest": "recallnest",
+  "openclaw": "openclaw", "open claw": "openclaw",
+  "ultramemory": "ultramemory", "ultra memory": "ultramemory",
+  "cc": "claude code", "claude code": "claude code",
+  "mcp": "model context protocol",
+};
+
+// ---------------------------------------------------------------------------
+// Resolver class
+// ---------------------------------------------------------------------------
+
+export class EntityResolver {
+  private aliases: Map<string, string>;
+
+  constructor(
+    builtinAliases: Record<string, string> = BUILTIN_ALIASES,
+    userAliases?: Record<string, string>,
+  ) {
+    this.aliases = new Map();
+    for (const [k, v] of Object.entries(builtinAliases)) {
+      this.aliases.set(k.toLowerCase(), v);
+    }
+    if (userAliases) {
+      for (const [k, v] of Object.entries(userAliases)) {
+        this.aliases.set(k.toLowerCase(), v);
+      }
+    }
+  }
+
+  /** Resolve a single term to its canonical form. Returns the term as-is if no alias found. */
+  resolve(term: string): string {
+    return this.aliases.get(term.toLowerCase().trim()) ?? term;
+  }
+
+  /** Resolve all recognizable entities in a text string. Returns text with aliases replaced. */
+  resolveText(text: string): string {
+    let result = text;
+    // Sort aliases by length descending to match longer aliases first
+    const sorted = [...this.aliases.entries()].sort((a, b) => b[0].length - a[0].length);
+    for (const [alias, canonical] of sorted) {
+      if (alias === canonical) continue; // skip identity mappings
+      const escaped = alias.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+      // Always use word boundary to prevent matching inside other words
+      // (e.g. "pt" inside "typescript")
+      const re = new RegExp(`\\b${escaped}\\b`, "gi");
+      result = result.replace(re, canonical);
+    }
+    return result;
+  }
+
+  /** Check if a term has a known alias. */
+  hasAlias(term: string): boolean {
+    return this.aliases.has(term.toLowerCase().trim());
+  }
+
+  /** Get all known aliases. */
+  get size(): number {
+    return this.aliases.size;
+  }
+}
+
+/** Default singleton resolver with built-in aliases. */
+export const defaultResolver = new EntityResolver();
+
+/**
+ * Resolve entities in a query string using the default resolver.
+ * Intended to be called before query expansion in the search pipeline.
+ */
+export function resolveQueryEntities(query: string): string {
+  return defaultResolver.resolveText(query);
+}