@cmdoss/memwal-sdk 0.9.0 → 1.0.0

package/src/services/EmbeddingService.ts

@@ -1,796 +1,796 @@
-/**
- * EmbeddingService - AI SDK Integration
- *
- * Refactored to use Vercel AI SDK as the underlying embedding provider.
- * Supports any AI SDK compatible provider (OpenAI, Google, Cohere, etc.)
- * while maintaining backward compatibility with existing PDW code.
- *
- * OpenRouter now uses the official @openrouter/sdk instead of raw fetch calls.
- *
- * Key features:
- * - Provider-agnostic: Accept any ai-sdk EmbeddingModel
- * - Backward compatible: Existing code continues to work
- * - Flexible configuration: Direct model OR provider config
- */
-
-import type { EmbeddingModelV2 } from '@ai-sdk/provider';
-import { embed, embedMany } from 'ai';
-import { createGoogleGenerativeAI } from '@ai-sdk/google';
-import { createOpenAI } from '@ai-sdk/openai';
-import { OpenRouter } from '@openrouter/sdk';
-
-// Type alias for embedding models - V2 is the default in AI SDK v5
-type EmbeddingModel<VALUE> = EmbeddingModelV2<VALUE>;
-
-// Provider instances (lazily initialized)
-let googleProvider: ReturnType<typeof createGoogleGenerativeAI> | null = null;
-let openaiProvider: ReturnType<typeof createOpenAI> | null = null;
-let openrouterProvider: ReturnType<typeof createOpenAI> | null = null;
-let cohereProvider: any = null;
-
-export interface EmbeddingConfig {
-  /**
-   * Option 1: Direct ai-sdk model (most flexible)
-   * User provides their own EmbeddingModel from any provider
-   *
-   * For backward compatibility, also accepts string (treated as modelName)
-   *
-   * @example
-   * ```typescript
-   * import { openai } from '@ai-sdk/openai';
-   * const service = new EmbeddingService({
-   *   model: openai.embedding('text-embedding-3-large')
-   * });
-   *
-   * // Backward compatible:
-   * const service = new EmbeddingService({
-   *   model: 'text-embedding-004', // Treated as modelName
-   *   apiKey: 'your-key'
-   * });
-   * ```
-   */
-  model?: EmbeddingModel<string> | string;
-
-  /**
-   * Option 2: Provider-based configuration
-   * PDW creates the model from provider settings
-   *
-   * - google: Direct Google AI API
-   * - openai: Direct OpenAI API
-   * - openrouter: OpenRouter API gateway (supports multiple models)
-   * - cohere: Direct Cohere API
-   */
-  provider?: 'google' | 'openai' | 'openrouter' | 'cohere';
-
-  /**
-   * API key for the provider
-   * Falls back to environment variables:
-   * - GEMINI_API_KEY or GOOGLE_AI_API_KEY (for google)
-   * - OPENAI_API_KEY (for openai)
-   * - OPENROUTER_API_KEY (for openrouter)
-   * - COHERE_API_KEY (for cohere)
-   */
-  apiKey?: string;
-
-  /**
-   * Model name to use
-   * - Google: 'text-embedding-004', 'gemini-embedding-001'
-   * - OpenAI: 'text-embedding-3-small', 'text-embedding-3-large'
-   * - OpenRouter: 'google/gemini-embedding-001', 'openai/text-embedding-3-small', etc.
-   * - Cohere: 'embed-english-v3.0', 'embed-multilingual-v3.0'
-   */
-  modelName?: string;
-
-  /**
-   * Embedding dimensions (optional, provider-dependent)
-   * - Google: Up to 3072
-   * - OpenAI: 256, 512, 1024, 1536, 3072 (depending on model)
-   * - OpenRouter: Depends on the underlying model
-   * - Cohere: Model-specific
-   */
-  dimensions?: number;
-
-  /**
-   * Rate limiting
-   */
-  requestsPerMinute?: number;
-}
-
-export interface EmbeddingOptions {
-  text: string;
-  type?: 'content' | 'metadata' | 'query';
-  taskType?: 'RETRIEVAL_QUERY' | 'RETRIEVAL_DOCUMENT' | 'SEMANTIC_SIMILARITY';
-}
-
-export interface EmbeddingResult {
-  vector: number[];
-  dimension: number;
-  model: string;
-  processingTime: number;
-  tokenCount?: number;
-}
-
-export interface BatchEmbeddingResult {
-  vectors: number[][];
-  dimension: number;
-  model: string;
-  totalProcessingTime: number;
-  averageProcessingTime: number;
-  successCount: number;
-  failedCount: number;
-}
-
-/**
- * Embedding service using Vercel AI SDK
- * Supports all AI SDK compatible providers
- * OpenRouter uses the official @openrouter/sdk for better type safety
- */
-export class EmbeddingService {
-  private embeddingModel: EmbeddingModel<string> | null = null;
-  private modelName: string;
-  private dimensions: number;
-  private requestCount = 0;
-  private lastReset = Date.now();
-  private readonly maxRequestsPerMinute: number;
-  private provider: 'google' | 'openai' | 'openrouter' | 'cohere' | 'custom';
-  private apiKey: string = '';
-  private openRouterClient: OpenRouter | null = null;
-
-  constructor(config: EmbeddingConfig = {}) {
-    this.maxRequestsPerMinute = config.requestsPerMinute || 1500;
-
-    // Case 1: Direct model provided (most flexible)
-    if (config.model) {
-      // Backward compatibility: If model is a string, treat as modelName
-      if (typeof config.model === 'string') {
-        const modelNameFromString = config.model;
-        console.log(`🔄 Backward compatibility: treating model string "${modelNameFromString}" as modelName`);
-
-        // Treat string as modelName and use provider config path
-        const provider = config.provider || 'google';
-        this.apiKey = this.resolveApiKey(provider, config.apiKey);
-
-        if (!this.apiKey) {
-          throw new Error(
-            `API key is required for ${provider} provider. ` +
-            `Provide it via config.apiKey or environment variable.`
-          );
-        }
-
-        this.provider = provider;
-        this.modelName = modelNameFromString;
-        this.dimensions = config.dimensions || this.getDefaultDimensions(provider);
-
-        // OpenRouter uses SDK, others use AI SDK
-        if (provider === 'openrouter') {
-          this.openRouterClient = new OpenRouter({ apiKey: this.apiKey });
-        } else {
-          this.embeddingModel = this.createModel(provider, this.apiKey, this.modelName);
-        }
-
-        console.log(`✅ EmbeddingService initialized with ${provider} provider (${this.modelName}) [backward compat mode]`);
-        return;
-      }
-
-      // New behavior: Direct EmbeddingModel from ai-sdk
-      this.embeddingModel = config.model;
-      this.modelName = 'custom';
-      this.dimensions = config.dimensions || 768; // Default 768 for speed (was 3072)
-      this.provider = 'custom';
-      console.log('✅ EmbeddingService initialized with custom ai-sdk model');
-      return;
-    }
-
-    // Case 2: Provider-based configuration
-    const provider = config.provider || 'google'; // Default to google for backward compat
-    this.apiKey = this.resolveApiKey(provider, config.apiKey);
-
-    if (!this.apiKey) {
-      throw new Error(
-        `API key is required for ${provider} provider. ` +
-        `Provide it via config.apiKey or environment variable.`
-      );
-    }
-
-    this.provider = provider;
-    this.modelName = config.modelName || this.getDefaultModelName(provider);
-    this.dimensions = config.dimensions || this.getDefaultDimensions(provider);
-
-    // OpenRouter uses SDK, others use AI SDK
-    if (provider === 'openrouter') {
-      this.openRouterClient = new OpenRouter({ apiKey: this.apiKey });
-    } else {
-      this.embeddingModel = this.createModel(provider, this.apiKey, this.modelName);
-    }
-
-    console.log(`✅ EmbeddingService initialized with ${provider} provider (${this.modelName})`);
-  }
-
-  /**
-   * Resolve API key from config or environment
-   */
-  private resolveApiKey(provider: string, configKey?: string): string {
-    if (configKey) return configKey;
-
-    switch (provider) {
-      case 'google':
-        return process.env.GEMINI_API_KEY || process.env.GOOGLE_AI_API_KEY || '';
-      case 'openai':
-        return process.env.OPENAI_API_KEY || '';
-      case 'openrouter':
-        return process.env.OPENROUTER_API_KEY || '';
-      case 'cohere':
-        return process.env.COHERE_API_KEY || '';
-      default:
-        return '';
-    }
-  }
-
-  /**
-   * Get default model name for provider
-   */
-  private getDefaultModelName(provider: string): string {
-    switch (provider) {
-      case 'google':
-        return 'text-embedding-004';
-      case 'openai':
-        return 'text-embedding-3-small';
-      case 'openrouter':
-        return 'google/gemini-embedding-001'; // Default OpenRouter embedding model
-      case 'cohere':
-        return 'embed-english-v3.0';
-      default:
-        return 'text-embedding-004';
-    }
-  }
-
-  /**
-   * Get default dimensions for provider
-   *
-   * Default is now 768 for faster performance:
-   * - 4x smaller vectors = faster indexing & search
-   * - ~4x less storage space
-   * - Minimal quality loss for most use cases
-   *
-   * Users can override via config.dimensions
-   */
-  private getDefaultDimensions(provider: string): number {
-    switch (provider) {
-      case 'google':
-        return 768; // text-embedding-004 supports output_dimensionality
-      case 'openai':
-        return 768; // text-embedding-3-small supports dimensions param
-      case 'openrouter':
-        return 768; // Most models support dimension truncation
-      case 'cohere':
-        return 768; // embed-english-v3.0 supports dimensions
-      default:
-        return 768; // Default to 768 for speed
-    }
-  }
-
-  /**
-   * Create embedding model from provider
-   */
-  private createModel(
-    provider: string,
-    apiKey: string,
-    modelName: string
-  ): EmbeddingModel<string> {
-    switch (provider) {
-      case 'google': {
-        if (!googleProvider) {
-          googleProvider = createGoogleGenerativeAI({ apiKey });
-        }
-        return googleProvider.textEmbeddingModel(modelName);
-      }
-
-      case 'openai': {
-        if (!openaiProvider) {
-          openaiProvider = createOpenAI({ apiKey });
-        }
-        // OpenAI returns EmbeddingModelV2 but is compatible with ai SDK
-        return openaiProvider.textEmbeddingModel(modelName) as unknown as EmbeddingModel<string>;
-      }
-
-      case 'openrouter': {
-        // OpenRouter uses OpenAI-compatible API with custom baseURL
-        if (!openrouterProvider) {
-          openrouterProvider = createOpenAI({
-            baseURL: 'https://openrouter.ai/api/v1',
-            apiKey,
-          });
-        }
-        // OpenRouter embedding models use the same interface as OpenAI
-        return openrouterProvider.textEmbeddingModel(modelName) as unknown as EmbeddingModel<string>;
-      }
-
-      case 'cohere': {
-        if (!cohereProvider) {
-          throw new Error(
-            'Cohere provider requires manual initialization. ' +
-            'Import createCohere from @ai-sdk/cohere and set cohereProvider before use.'
-          );
-        }
-        return cohereProvider.textEmbedding(modelName);
-      }
-
-      default:
-        throw new Error(`Unsupported provider: ${provider}`);
-    }
-  }
-
-  /**
-   * Generate embedding for a single text
-   */
-  async embedText(options: EmbeddingOptions): Promise<EmbeddingResult> {
-    const startTime = Date.now();
-
-    // Validate input
-    if (!options.text || typeof options.text !== 'string' || options.text.trim().length === 0) {
-      throw new Error('Invalid or empty text provided for embedding');
-    }
-
-    await this.checkRateLimit();
-
-    try {
-      // OpenRouter uses native fetch API for better compatibility
-      if (this.provider === 'openrouter') {
-        return await this.embedTextOpenRouter(options.text, startTime);
-      }
-
-      // Other providers use AI SDK
-      if (!this.embeddingModel) {
-        throw new Error('Embedding model not initialized');
-      }
-
-      const result = await embed({
-        model: this.embeddingModel,
-        value: options.text,
-        ...this.getProviderOptions(options),
-      });
-
-      this.requestCount++;
-
-      return {
-        vector: result.embedding,
-        dimension: result.embedding.length,
-        model: this.modelName,
-        processingTime: Date.now() - startTime,
-        tokenCount: result.usage?.tokens,
-      };
-    } catch (error) {
-      throw new Error(
-        `Embedding generation failed: ${error instanceof Error ? error.message : 'Unknown error'}`
-      );
-    }
-  }
-
-  /**
-   * Generate embedding using OpenRouter SDK
-   * Uses official @openrouter/sdk for embeddings
-   * Passes dimensions parameter to truncate output vectors
-   */
-  private async embedTextOpenRouter(text: string, startTime: number): Promise<EmbeddingResult> {
-    if (!this.openRouterClient) {
-      throw new Error('OpenRouter client not initialized');
-    }
-
-    // Build request with optional dimensions parameter
-    const request: any = {
-      model: this.modelName,
-      input: text
-    };
-
-    // Add dimensions if configured (enables output truncation)
-    if (this.dimensions && this.dimensions < 3072) {
-      request.dimensions = this.dimensions;
-    }
-
-    const result = await this.openRouterClient.embeddings.generate(request);
-
-    // Handle union type - result can be string or object
-    if (typeof result === 'string') {
-      throw new Error('Unexpected string response from OpenRouter embeddings API');
-    }
-
-    const data = (result as any).data;
-    if (!data || !data[0] || !data[0].embedding) {
-      throw new Error('Invalid response from OpenRouter embeddings API');
-    }
-
-    this.requestCount++;
-
-    // Handle embedding which can be string or number[]
-    const embedding = data[0].embedding;
-    const vector = typeof embedding === 'string'
-      ? JSON.parse(embedding) as number[]
-      : embedding as number[];
-
-    const usage = (result as any).usage;
-
-    return {
-      vector,
-      dimension: vector.length,
-      model: this.modelName,
-      processingTime: Date.now() - startTime,
-      tokenCount: usage?.totalTokens
-    };
-  }
-
-  /**
-   * Generate embeddings for multiple texts (batched)
-   */
-  async embedBatch(
-    texts: string[],
-    options: Omit<EmbeddingOptions, 'text'> = {}
-  ): Promise<BatchEmbeddingResult> {
-    const startTime = Date.now();
-    let successCount = 0;
-    let failedCount = 0;
-
-    try {
-      await this.checkRateLimit();
-
-      // OpenRouter uses native fetch API for better compatibility
-      if (this.provider === 'openrouter') {
-        return await this.embedBatchOpenRouter(texts, startTime);
-      }
-
-      // Other providers use AI SDK
-      if (!this.embeddingModel) {
-        throw new Error('Embedding model not initialized');
-      }
-
-      const result = await embedMany({
-        model: this.embeddingModel,
-        values: texts,
-        ...this.getProviderOptions(options as EmbeddingOptions),
-      });
-
-      successCount = result.embeddings.length;
-      const totalTime = Date.now() - startTime;
-
-      return {
-        vectors: result.embeddings,
-        dimension: result.embeddings[0]?.length || this.dimensions,
-        model: this.modelName,
-        totalProcessingTime: totalTime,
-        averageProcessingTime: totalTime / texts.length,
-        successCount,
-        failedCount,
-      };
-    } catch (error) {
-      throw new Error(
-        `Batch embedding failed: ${error instanceof Error ? error.message : 'Unknown error'}`
-      );
-    }
-  }
-
-  /**
-   * Generate batch embeddings using OpenRouter SDK
-   * Passes dimensions parameter to truncate output vectors
-   */
-  private async embedBatchOpenRouter(texts: string[], startTime: number): Promise<BatchEmbeddingResult> {
-    if (!this.openRouterClient) {
-      throw new Error('OpenRouter client not initialized');
-    }
-
-    // Build request with optional dimensions parameter
-    const request: any = {
-      model: this.modelName,
-      input: texts
-    };
-
-    // Add dimensions if configured (enables output truncation)
-    if (this.dimensions && this.dimensions < 3072) {
-      request.dimensions = this.dimensions;
-    }
-
-    const result = await this.openRouterClient.embeddings.generate(request);
-
-    // Handle union type - result can be string or object
-    if (typeof result === 'string') {
-      throw new Error('Unexpected string response from OpenRouter embeddings API');
-    }
-
-    const data = (result as any).data;
-    if (!data || !Array.isArray(data)) {
-      throw new Error('Invalid response from OpenRouter embeddings API');
-    }
-
-    // Sort by index to ensure correct order
-    const sortedData = [...data].sort((a: any, b: any) => (a.index || 0) - (b.index || 0));
-    const vectors = sortedData.map((item: any) => {
-      const embedding = item.embedding;
-      return typeof embedding === 'string'
-        ? JSON.parse(embedding) as number[]
-        : embedding as number[];
-    });
-
-    this.requestCount++;
-    const totalTime = Date.now() - startTime;
-
-    return {
-      vectors,
-      dimension: vectors[0]?.length || this.dimensions,
-      model: this.modelName,
-      totalProcessingTime: totalTime,
-      averageProcessingTime: totalTime / texts.length,
-      successCount: vectors.length,
-      failedCount: texts.length - vectors.length
-    };
-  }
-
-  /**
-   * Get provider-specific options
-   */
-  private getProviderOptions(options: EmbeddingOptions): any {
-    const providerOpts: any = {};
-
-    if (this.provider === 'google') {
-      providerOpts.providerOptions = {
-        google: {
-          outputDimensionality: this.dimensions,
-          taskType: this.getGoogleTaskType(options.type),
-        },
-      };
-    } else if (this.provider === 'openai') {
-      providerOpts.providerOptions = {
-        openai: {
-          dimensions: this.dimensions,
-        },
-      };
-    } else if (this.provider === 'openrouter') {
-      // OpenRouter uses OpenAI-compatible options
-      // Note: dimensions may not be supported for all models via OpenRouter
-      providerOpts.providerOptions = {
-        openai: {
-          dimensions: this.dimensions,
-        },
-      };
-    } else if (this.provider === 'cohere') {
-      providerOpts.providerOptions = {
-        cohere: {
-          inputType: this.getCohereInputType(options.type),
-        },
-      };
-    }
-
-    return providerOpts;
-  }
-
-  /**
-   * Map PDW type to Google task type
-   */
-  private getGoogleTaskType(type?: string): string {
-    switch (type) {
-      case 'query':
-        return 'RETRIEVAL_QUERY';
-      case 'content':
-        return 'RETRIEVAL_DOCUMENT';
-      case 'metadata':
-        return 'SEMANTIC_SIMILARITY';
-      default:
-        return 'RETRIEVAL_DOCUMENT';
-    }
-  }
-
-  /**
-   * Map PDW type to Cohere input type
-   */
-  private getCohereInputType(type?: string): string {
-    switch (type) {
-      case 'query':
-        return 'search_query';
-      case 'content':
-        return 'search_document';
-      default:
-        return 'search_document';
-    }
-  }
-
-  /**
-   * Calculate cosine similarity between two vectors
-   */
-  calculateCosineSimilarity(vectorA: number[], vectorB: number[]): number {
-    if (vectorA.length !== vectorB.length) {
-      throw new Error(`Vector dimension mismatch: ${vectorA.length} vs ${vectorB.length}`);
-    }
-
-    let dotProduct = 0;
-    let normA = 0;
-    let normB = 0;
-
-    for (let i = 0; i < vectorA.length; i++) {
-      dotProduct += vectorA[i] * vectorB[i];
-      normA += vectorA[i] * vectorA[i];
-      normB += vectorB[i] * vectorB[i];
-    }
-
-    const magnitude = Math.sqrt(normA) * Math.sqrt(normB);
-
-    if (magnitude === 0) {
-      return 0;
-    }
-
-    return dotProduct / magnitude;
-  }
-
-  /**
-   * Calculate Euclidean distance between two vectors
-   */
-  calculateEuclideanDistance(vectorA: number[], vectorB: number[]): number {
-    if (vectorA.length !== vectorB.length) {
-      throw new Error(`Vector dimension mismatch: ${vectorA.length} vs ${vectorB.length}`);
-    }
-
-    let sum = 0;
-    for (let i = 0; i < vectorA.length; i++) {
-      const diff = vectorA[i] - vectorB[i];
-      sum += diff * diff;
-    }
-
-    return Math.sqrt(sum);
-  }
-
-  /**
-   * Normalize a vector to unit length
-   */
-  normalizeVector(vector: number[]): number[] {
-    const magnitude = Math.sqrt(vector.reduce((sum, val) => sum + val * val, 0));
-
-    if (magnitude === 0) {
-      return vector;
-    }
-
-    return vector.map(val => val / magnitude);
-  }
-
-  /**
-   * Find the most similar vectors to a query vector
-   */
-  findMostSimilar(
-    queryVector: number[],
-    candidateVectors: number[][],
-    k: number = 5
-  ): Array<{ index: number; similarity: number; distance: number }> {
-    const similarities = candidateVectors.map((vector, index) => {
-      const similarity = this.calculateCosineSimilarity(queryVector, vector);
-      const distance = this.calculateEuclideanDistance(queryVector, vector);
-
-      return { index, similarity, distance };
-    });
-
-    similarities.sort((a, b) => b.similarity - a.similarity);
-
-    return similarities.slice(0, k);
-  }
-
-  /**
-   * Get embedding statistics
-   */
-  getStats(): {
-    totalRequests: number;
-    requestsThisMinute: number;
-    model: string;
-    dimensions: number;
-    rateLimit: number;
-    provider: string;
-  } {
-    const now = Date.now();
-    const requestsThisMinute = (now - this.lastReset) < 60000 ? this.requestCount : 0;
-
-    return {
-      totalRequests: this.requestCount,
-      requestsThisMinute,
-      model: this.modelName,
-      dimensions: this.dimensions,
-      rateLimit: this.maxRequestsPerMinute,
-      provider: this.provider,
-    };
-  }
-
-  /**
-   * Reset rate limiting counters
-   */
-  private resetRateLimit(): void {
-    const now = Date.now();
-    if (now - this.lastReset >= 60000) {
-      this.requestCount = 0;
-      this.lastReset = now;
-    }
-  }
-
-  /**
-   * Check rate limiting and wait if necessary
-   */
-  private async checkRateLimit(): Promise<void> {
-    this.resetRateLimit();
-
-    if (this.requestCount >= this.maxRequestsPerMinute) {
-      const waitTime = 60000 - (Date.now() - this.lastReset);
-      if (waitTime > 0) {
-        if (process.env.NODE_ENV === 'development') {
-          console.warn(`Rate limit reached, waiting ${waitTime}ms`);
-        }
-        await this.delay(waitTime);
-        this.resetRateLimit();
-      }
-    }
-  }
-
-  /**
-   * Utility delay function
-   */
-  private delay(ms: number): Promise<void> {
-    return new Promise(resolve => setTimeout(resolve, ms));
-  }
-}
-
-export default EmbeddingService;
-
-// ==================== Singleton Pattern ====================
-
-/**
- * Generate config key for singleton cache
- */
-function getConfigKey(config: EmbeddingConfig): string {
-  const provider = config.provider || 'google';
-  const modelName = typeof config.model === 'string'
-    ? config.model
-    : (config.modelName || 'default');
-  const dimensions = config.dimensions || 'default';
-  return `${provider}:${modelName}:${dimensions}`;
-}
-
-/** Singleton cache */
-const sharedInstances = new Map<string, EmbeddingService>();
-
-/**
- * Get or create a shared EmbeddingService instance (Singleton)
- *
- * All clients with same provider/model/dimensions share one instance.
- * Reduces memory usage and connection overhead.
- *
- * @example
- * ```typescript
- * // Instead of: new EmbeddingService({ apiKey, modelName })
- * const embedding = getSharedEmbeddingService({ apiKey, modelName });
- * ```
- */
-export function getSharedEmbeddingService(config: EmbeddingConfig): EmbeddingService {
-  const key = getConfigKey(config);
-
-  let instance = sharedInstances.get(key);
-  if (!instance) {
-    console.log(`🔧 [Singleton] Creating shared EmbeddingService: ${key}`);
-    instance = new EmbeddingService(config);
-    sharedInstances.set(key, instance);
-  }
-
-  return instance;
-}
-
-/**
- * Clear all shared instances (for testing)
- */
-export function clearSharedEmbeddingServices(): void {
-  sharedInstances.clear();
-}
-
-/**
- * Get singleton stats
- */
-export function getSharedEmbeddingStats(): {
-  instanceCount: number;
-  instances: Array<{ key: string; stats: ReturnType<EmbeddingService['getStats']> }>;
-} {
-  return {
-    instanceCount: sharedInstances.size,
-    instances: Array.from(sharedInstances.entries()).map(([key, svc]) => ({
-      key,
-      stats: svc.getStats(),
-    })),
-  };
-}
+/**
+ * EmbeddingService - AI SDK Integration
+ *
+ * Refactored to use Vercel AI SDK as the underlying embedding provider.
+ * Supports any AI SDK compatible provider (OpenAI, Google, Cohere, etc.)
+ * while maintaining backward compatibility with existing PDW code.
+ *
+ * OpenRouter now uses the official @openrouter/sdk instead of raw fetch calls.
+ *
+ * Key features:
+ * - Provider-agnostic: Accept any ai-sdk EmbeddingModel
+ * - Backward compatible: Existing code continues to work
+ * - Flexible configuration: Direct model OR provider config
+ */
+
+import type { EmbeddingModelV2 } from '@ai-sdk/provider';
+import { embed, embedMany } from 'ai';
+import { createGoogleGenerativeAI } from '@ai-sdk/google';
+import { createOpenAI } from '@ai-sdk/openai';
+import { OpenRouter } from '@openrouter/sdk';
+
+// Type alias for embedding models - V2 is the default in AI SDK v5
+type EmbeddingModel<VALUE> = EmbeddingModelV2<VALUE>;
+
+// Provider instances (lazily initialized)
+let googleProvider: ReturnType<typeof createGoogleGenerativeAI> | null = null;
+let openaiProvider: ReturnType<typeof createOpenAI> | null = null;
+let openrouterProvider: ReturnType<typeof createOpenAI> | null = null;
+let cohereProvider: any = null;
+
+export interface EmbeddingConfig {
+  /**
+   * Option 1: Direct ai-sdk model (most flexible)
+   * User provides their own EmbeddingModel from any provider
+   *
+   * For backward compatibility, also accepts string (treated as modelName)
+   *
+   * @example
+   * ```typescript
+   * import { openai } from '@ai-sdk/openai';
+   * const service = new EmbeddingService({
+   *   model: openai.embedding('text-embedding-3-large')
+   * });
+   *
+   * // Backward compatible:
+   * const service = new EmbeddingService({
+   *   model: 'text-embedding-004', // Treated as modelName
+   *   apiKey: 'your-key'
+   * });
+   * ```
+   */
+  model?: EmbeddingModel<string> | string;
+
+  /**
+   * Option 2: Provider-based configuration
+   * PDW creates the model from provider settings
+   *
+   * - google: Direct Google AI API
+   * - openai: Direct OpenAI API
+   * - openrouter: OpenRouter API gateway (supports multiple models)
+   * - cohere: Direct Cohere API
+   */
+  provider?: 'google' | 'openai' | 'openrouter' | 'cohere';
+
+  /**
+   * API key for the provider
+   * Falls back to environment variables:
+   * - GEMINI_API_KEY or GOOGLE_AI_API_KEY (for google)
+   * - OPENAI_API_KEY (for openai)
+   * - OPENROUTER_API_KEY (for openrouter)
+   * - COHERE_API_KEY (for cohere)
+   */
+  apiKey?: string;
+
+  /**
+   * Model name to use
+   * - Google: 'text-embedding-004', 'gemini-embedding-001'
+   * - OpenAI: 'text-embedding-3-small', 'text-embedding-3-large'
+   * - OpenRouter: 'google/gemini-embedding-001', 'openai/text-embedding-3-small', etc.
+   * - Cohere: 'embed-english-v3.0', 'embed-multilingual-v3.0'
+   */
+  modelName?: string;
+
+  /**
+   * Embedding dimensions (optional, provider-dependent)
+   * - Google: Up to 3072
+   * - OpenAI: 256, 512, 1024, 1536, 3072 (depending on model)
+   * - OpenRouter: Depends on the underlying model
+   * - Cohere: Model-specific
+   */
+  dimensions?: number;
+
+  /**
+   * Rate limiting
+   */
+  requestsPerMinute?: number;
+}
+
+export interface EmbeddingOptions {
+  text: string;
+  type?: 'content' | 'metadata' | 'query';
+  taskType?: 'RETRIEVAL_QUERY' | 'RETRIEVAL_DOCUMENT' | 'SEMANTIC_SIMILARITY';
+}
+
+export interface EmbeddingResult {
+  vector: number[];
+  dimension: number;
+  model: string;
+  processingTime: number;
+  tokenCount?: number;
+}
+
+export interface BatchEmbeddingResult {
+  vectors: number[][];
+  dimension: number;
+  model: string;
+  totalProcessingTime: number;
+  averageProcessingTime: number;
+  successCount: number;
+  failedCount: number;
+}
+
+/**
+ * Embedding service using Vercel AI SDK
+ * Supports all AI SDK compatible providers
+ * OpenRouter uses the official @openrouter/sdk for better type safety
+ */
+export class EmbeddingService {
+  private embeddingModel: EmbeddingModel<string> | null = null;
+  private modelName: string;
+  private dimensions: number;
+  private requestCount = 0;
+  private lastReset = Date.now();
+  private readonly maxRequestsPerMinute: number;
+  private provider: 'google' | 'openai' | 'openrouter' | 'cohere' | 'custom';
+  private apiKey: string = '';
+  private openRouterClient: OpenRouter | null = null;
+
+  constructor(config: EmbeddingConfig = {}) {
+    this.maxRequestsPerMinute = config.requestsPerMinute || 1500;
+
+    // Case 1: Direct model provided (most flexible)
+    if (config.model) {
+      // Backward compatibility: If model is a string, treat as modelName
+      if (typeof config.model === 'string') {
+        const modelNameFromString = config.model;
+        console.log(`🔄 Backward compatibility: treating model string "${modelNameFromString}" as modelName`);
+
+        // Treat string as modelName and use provider config path
+        const provider = config.provider || 'google';
+        this.apiKey = this.resolveApiKey(provider, config.apiKey);
+
+        if (!this.apiKey) {
+          throw new Error(
+            `API key is required for ${provider} provider. ` +
+            `Provide it via config.apiKey or environment variable.`
+          );
+        }
+
+        this.provider = provider;
+        this.modelName = modelNameFromString;
+        this.dimensions = config.dimensions || this.getDefaultDimensions(provider);
+
+        // OpenRouter uses SDK, others use AI SDK
+        if (provider === 'openrouter') {
+          this.openRouterClient = new OpenRouter({ apiKey: this.apiKey });
+        } else {
+          this.embeddingModel = this.createModel(provider, this.apiKey, this.modelName);
+        }
+
+        console.log(`✅ EmbeddingService initialized with ${provider} provider (${this.modelName}) [backward compat mode]`);
+        return;
+      }
+
+      // New behavior: Direct EmbeddingModel from ai-sdk
+      this.embeddingModel = config.model;
+      this.modelName = 'custom';
+      this.dimensions = config.dimensions || 768; // Default 768 for speed (was 3072)
+      this.provider = 'custom';
+      console.log('✅ EmbeddingService initialized with custom ai-sdk model');
+      return;
+    }
+
+    // Case 2: Provider-based configuration
+    const provider = config.provider || 'google'; // Default to google for backward compat
+    this.apiKey = this.resolveApiKey(provider, config.apiKey);
+
+    if (!this.apiKey) {
+      throw new Error(
+        `API key is required for ${provider} provider. ` +
+        `Provide it via config.apiKey or environment variable.`
+      );
+    }
+
+    this.provider = provider;
+    this.modelName = config.modelName || this.getDefaultModelName(provider);
+    this.dimensions = config.dimensions || this.getDefaultDimensions(provider);
+
+    // OpenRouter uses SDK, others use AI SDK
+    if (provider === 'openrouter') {
+      this.openRouterClient = new OpenRouter({ apiKey: this.apiKey });
+    } else {
+      this.embeddingModel = this.createModel(provider, this.apiKey, this.modelName);
+    }
+
+    console.log(`✅ EmbeddingService initialized with ${provider} provider (${this.modelName})`);
+  }
+
+  /**
+   * Resolve API key from config or environment
+   */
+  private resolveApiKey(provider: string, configKey?: string): string {
+    if (configKey) return configKey;
+
+    switch (provider) {
+      case 'google':
+        return process.env.GEMINI_API_KEY || process.env.GOOGLE_AI_API_KEY || '';
+      case 'openai':
+        return process.env.OPENAI_API_KEY || '';
+      case 'openrouter':
+        return process.env.OPENROUTER_API_KEY || '';
+      case 'cohere':
+        return process.env.COHERE_API_KEY || '';
+      default:
+        return '';
+    }
+  }
+
+  /**
+   * Get default model name for provider
+   */
+  private getDefaultModelName(provider: string): string {
+    switch (provider) {
+      case 'google':
+        return 'text-embedding-004';
+      case 'openai':
+        return 'text-embedding-3-small';
+      case 'openrouter':
+        return 'google/gemini-embedding-001'; // Default OpenRouter embedding model
+      case 'cohere':
+        return 'embed-english-v3.0';
+      default:
+        return 'text-embedding-004';
+    }
+  }
+
+  /**
+   * Get default dimensions for provider
+   *
+   * Default is now 768 for faster performance:
+   * - 4x smaller vectors = faster indexing & search
+   * - ~4x less storage space
+   * - Minimal quality loss for most use cases
+   *
+   * Users can override via config.dimensions
+   */
+  private getDefaultDimensions(provider: string): number {
+    switch (provider) {
+      case 'google':
+        return 768; // text-embedding-004 supports output_dimensionality
+      case 'openai':
+        return 768; // text-embedding-3-small supports dimensions param
+      case 'openrouter':
+        return 768; // Most models support dimension truncation
+      case 'cohere':
+        return 768; // embed-english-v3.0 supports dimensions
+      default:
+        return 768; // Default to 768 for speed
+    }
+  }
+
+  /**
+   * Create embedding model from provider
+   */
+  private createModel(
+    provider: string,
+    apiKey: string,
+    modelName: string
+  ): EmbeddingModel<string> {
+    switch (provider) {
+      case 'google': {
+        if (!googleProvider) {
+          googleProvider = createGoogleGenerativeAI({ apiKey });
+        }
+        return googleProvider.textEmbeddingModel(modelName);
+      }
+
+      case 'openai': {
+        if (!openaiProvider) {
+          openaiProvider = createOpenAI({ apiKey });
+        }
+        // OpenAI returns EmbeddingModelV2 but is compatible with ai SDK
+        return openaiProvider.textEmbeddingModel(modelName) as unknown as EmbeddingModel<string>;
+      }
+
+      case 'openrouter': {
+        // OpenRouter uses OpenAI-compatible API with custom baseURL
+        if (!openrouterProvider) {
+          openrouterProvider = createOpenAI({
+            baseURL: 'https://openrouter.ai/api/v1',
+            apiKey,
+          });
+        }
+        // OpenRouter embedding models use the same interface as OpenAI
+        return openrouterProvider.textEmbeddingModel(modelName) as unknown as EmbeddingModel<string>;
+      }
+
+      case 'cohere': {
+        if (!cohereProvider) {
+          throw new Error(
+            'Cohere provider requires manual initialization. ' +
+            'Import createCohere from @ai-sdk/cohere and set cohereProvider before use.'
+          );
+        }
+        return cohereProvider.textEmbedding(modelName);
+      }
+
+      default:
+        throw new Error(`Unsupported provider: ${provider}`);
+    }
+  }
+
+  /**
+   * Generate embedding for a single text
+   */
+  async embedText(options: EmbeddingOptions): Promise<EmbeddingResult> {
+    const startTime = Date.now();
+
+    // Validate input
+    if (!options.text || typeof options.text !== 'string' || options.text.trim().length === 0) {
+      throw new Error('Invalid or empty text provided for embedding');
+    }
+
+    await this.checkRateLimit();
+
+    try {
+      // OpenRouter uses native fetch API for better compatibility
+      if (this.provider === 'openrouter') {
+        return await this.embedTextOpenRouter(options.text, startTime);
+      }
+
+      // Other providers use AI SDK
+      if (!this.embeddingModel) {
+        throw new Error('Embedding model not initialized');
+      }
+
+      const result = await embed({
+        model: this.embeddingModel,
+        value: options.text,
+        ...this.getProviderOptions(options),
+      });
+
+      this.requestCount++;
+
+      return {
+        vector: result.embedding,
+        dimension: result.embedding.length,
+        model: this.modelName,
+        processingTime: Date.now() - startTime,
+        tokenCount: result.usage?.tokens,
+      };
+    } catch (error) {
+      throw new Error(
+        `Embedding generation failed: ${error instanceof Error ? error.message : 'Unknown error'}`
+      );
+    }
+  }
+
+  /**
+   * Generate embedding using OpenRouter SDK
+   * Uses official @openrouter/sdk for embeddings
+   * Passes dimensions parameter to truncate output vectors
+   */
+  private async embedTextOpenRouter(text: string, startTime: number): Promise<EmbeddingResult> {
+    if (!this.openRouterClient) {
+      throw new Error('OpenRouter client not initialized');
+    }
+
+    // Build request with optional dimensions parameter
+    const request: any = {
+      model: this.modelName,
+      input: text
+    };
+
+    // Add dimensions if configured (enables output truncation)
+    if (this.dimensions && this.dimensions < 3072) {
+      request.dimensions = this.dimensions;
+    }
+
+    const result = await this.openRouterClient.embeddings.generate(request);
+
+    // Handle union type - result can be string or object
+    if (typeof result === 'string') {
+      throw new Error('Unexpected string response from OpenRouter embeddings API');
+    }
+
+    const data = (result as any).data;
+    if (!data || !data[0] || !data[0].embedding) {
+      throw new Error('Invalid response from OpenRouter embeddings API');
+    }
+
+    this.requestCount++;
+
+    // Handle embedding which can be string or number[]
+    const embedding = data[0].embedding;
+    const vector = typeof embedding === 'string'
+      ? JSON.parse(embedding) as number[]
+      : embedding as number[];
+
+    const usage = (result as any).usage;
+
+    return {
+      vector,
+      dimension: vector.length,
+      model: this.modelName,
+      processingTime: Date.now() - startTime,
+      tokenCount: usage?.totalTokens
+    };
+  }
+
+  /**
+   * Generate embeddings for multiple texts (batched)
+   */
+  async embedBatch(
+    texts: string[],
+    options: Omit<EmbeddingOptions, 'text'> = {}
+  ): Promise<BatchEmbeddingResult> {
+    const startTime = Date.now();
+    let successCount = 0;
+    let failedCount = 0;
+
+    try {
+      await this.checkRateLimit();
+
+      // OpenRouter uses native fetch API for better compatibility
+      if (this.provider === 'openrouter') {
+        return await this.embedBatchOpenRouter(texts, startTime);
+      }
+
+      // Other providers use AI SDK
+      if (!this.embeddingModel) {
+        throw new Error('Embedding model not initialized');
+      }
+
+      const result = await embedMany({
+        model: this.embeddingModel,
+        values: texts,
+        ...this.getProviderOptions(options as EmbeddingOptions),
+      });
+
+      successCount = result.embeddings.length;
+      const totalTime = Date.now() - startTime;
+
+      return {
+        vectors: result.embeddings,
+        dimension: result.embeddings[0]?.length || this.dimensions,
+        model: this.modelName,
+        totalProcessingTime: totalTime,
+        averageProcessingTime: totalTime / texts.length,
+        successCount,
+        failedCount,
+      };
+    } catch (error) {
+      throw new Error(
+        `Batch embedding failed: ${error instanceof Error ? error.message : 'Unknown error'}`
+      );
+    }
+  }
+
+  /**
+   * Generate batch embeddings using OpenRouter SDK
+   * Passes dimensions parameter to truncate output vectors
+   */
+  private async embedBatchOpenRouter(texts: string[], startTime: number): Promise<BatchEmbeddingResult> {
+    if (!this.openRouterClient) {
+      throw new Error('OpenRouter client not initialized');
+    }
+
+    // Build request with optional dimensions parameter
+    const request: any = {
+      model: this.modelName,
+      input: texts
+    };
+
+    // Add dimensions if configured (enables output truncation)
+    if (this.dimensions && this.dimensions < 3072) {
+      request.dimensions = this.dimensions;
+    }
+
+    const result = await this.openRouterClient.embeddings.generate(request);
+
+    // Handle union type - result can be string or object
+    if (typeof result === 'string') {
+      throw new Error('Unexpected string response from OpenRouter embeddings API');
+    }
+
+    const data = (result as any).data;
+    if (!data || !Array.isArray(data)) {
+      throw new Error('Invalid response from OpenRouter embeddings API');
+    }
+
+    // Sort by index to ensure correct order
+    const sortedData = [...data].sort((a: any, b: any) => (a.index || 0) - (b.index || 0));
+    const vectors = sortedData.map((item: any) => {
+      const embedding = item.embedding;
+      return typeof embedding === 'string'
+        ? JSON.parse(embedding) as number[]
+        : embedding as number[];
+    });
+
+    this.requestCount++;
+    const totalTime = Date.now() - startTime;
+
+    return {
+      vectors,
+      dimension: vectors[0]?.length || this.dimensions,
+      model: this.modelName,
+      totalProcessingTime: totalTime,
+      averageProcessingTime: totalTime / texts.length,
+      successCount: vectors.length,
+      failedCount: texts.length - vectors.length
+    };
+  }
+
+  /**
+   * Get provider-specific options
+   */
+  private getProviderOptions(options: EmbeddingOptions): any {
+    const providerOpts: any = {};
+
+    if (this.provider === 'google') {
+      providerOpts.providerOptions = {
+        google: {
+          outputDimensionality: this.dimensions,
+          taskType: this.getGoogleTaskType(options.type),
+        },
+      };
+    } else if (this.provider === 'openai') {
+      providerOpts.providerOptions = {
+        openai: {
+          dimensions: this.dimensions,
+        },
+      };
+    } else if (this.provider === 'openrouter') {
+      // OpenRouter uses OpenAI-compatible options
+      // Note: dimensions may not be supported for all models via OpenRouter
+      providerOpts.providerOptions = {
+        openai: {
+          dimensions: this.dimensions,
+        },
+      };
+    } else if (this.provider === 'cohere') {
+      providerOpts.providerOptions = {
+        cohere: {
+          inputType: this.getCohereInputType(options.type),
+        },
+      };
+    }
+
+    return providerOpts;
+  }
+
+  /**
+   * Map PDW type to Google task type
+   */
+  private getGoogleTaskType(type?: string): string {
+    switch (type) {
+      case 'query':
+        return 'RETRIEVAL_QUERY';
+      case 'content':
+        return 'RETRIEVAL_DOCUMENT';
+      case 'metadata':
+        return 'SEMANTIC_SIMILARITY';
+      default:
+        return 'RETRIEVAL_DOCUMENT';
+    }
+  }
+
+  /**
+   * Map PDW type to Cohere input type
+   */
+  private getCohereInputType(type?: string): string {
+    switch (type) {
+      case 'query':
+        return 'search_query';
+      case 'content':
+        return 'search_document';
+      default:
+        return 'search_document';
+    }
+  }
+
+  /**
+   * Calculate cosine similarity between two vectors
+   */
+  calculateCosineSimilarity(vectorA: number[], vectorB: number[]): number {
+    if (vectorA.length !== vectorB.length) {
+      throw new Error(`Vector dimension mismatch: ${vectorA.length} vs ${vectorB.length}`);
+    }
+
+    let dotProduct = 0;
+    let normA = 0;
+    let normB = 0;
+
+    for (let i = 0; i < vectorA.length; i++) {
+      dotProduct += vectorA[i] * vectorB[i];
+      normA += vectorA[i] * vectorA[i];
+      normB += vectorB[i] * vectorB[i];
+    }
+
+    const magnitude = Math.sqrt(normA) * Math.sqrt(normB);
+
+    if (magnitude === 0) {
+      return 0;
+    }
+
+    return dotProduct / magnitude;
+  }
+
+  /**
+   * Calculate Euclidean distance between two vectors
+   */
+  calculateEuclideanDistance(vectorA: number[], vectorB: number[]): number {
+    if (vectorA.length !== vectorB.length) {
+      throw new Error(`Vector dimension mismatch: ${vectorA.length} vs ${vectorB.length}`);
+    }
+
+    let sum = 0;
+    for (let i = 0; i < vectorA.length; i++) {
+      const diff = vectorA[i] - vectorB[i];
+      sum += diff * diff;
+    }
+
+    return Math.sqrt(sum);
+  }
+
+  /**
+   * Normalize a vector to unit length
+   */
+  normalizeVector(vector: number[]): number[] {
+    const magnitude = Math.sqrt(vector.reduce((sum, val) => sum + val * val, 0));
+
+    if (magnitude === 0) {
+      return vector;
+    }
+
+    return vector.map(val => val / magnitude);
+  }
+
+  /**
+   * Find the most similar vectors to a query vector
+   */
+  findMostSimilar(
+    queryVector: number[],
+    candidateVectors: number[][],
+    k: number = 5
+  ): Array<{ index: number; similarity: number; distance: number }> {
+    const similarities = candidateVectors.map((vector, index) => {
+      const similarity = this.calculateCosineSimilarity(queryVector, vector);
+      const distance = this.calculateEuclideanDistance(queryVector, vector);
+
+      return { index, similarity, distance };
+    });
+
+    similarities.sort((a, b) => b.similarity - a.similarity);
+
+    return similarities.slice(0, k);
+  }
+
+  /**
+   * Get embedding statistics
+   */
+  getStats(): {
+    totalRequests: number;
+    requestsThisMinute: number;
+    model: string;
+    dimensions: number;
+    rateLimit: number;
+    provider: string;
+  } {
+    const now = Date.now();
+    const requestsThisMinute = (now - this.lastReset) < 60000 ? this.requestCount : 0;
+
+    return {
+      totalRequests: this.requestCount,
+      requestsThisMinute,
+      model: this.modelName,
+      dimensions: this.dimensions,
+      rateLimit: this.maxRequestsPerMinute,
+      provider: this.provider,
+    };
+  }
+
+  /**
+   * Reset rate limiting counters
+   */
+  private resetRateLimit(): void {
+    const now = Date.now();
+    if (now - this.lastReset >= 60000) {
+      this.requestCount = 0;
+      this.lastReset = now;
+    }
+  }
+
+  /**
+   * Check rate limiting and wait if necessary
+   */
+  private async checkRateLimit(): Promise<void> {
+    this.resetRateLimit();
+
+    if (this.requestCount >= this.maxRequestsPerMinute) {
+      const waitTime = 60000 - (Date.now() - this.lastReset);
+      if (waitTime > 0) {
+        if (process.env.NODE_ENV === 'development') {
+          console.warn(`Rate limit reached, waiting ${waitTime}ms`);
+        }
+        await this.delay(waitTime);
+        this.resetRateLimit();
+      }
+    }
+  }
+
+  /**
+   * Utility delay function
+   */
+  private delay(ms: number): Promise<void> {
+    return new Promise(resolve => setTimeout(resolve, ms));
+  }
+}
+
+export default EmbeddingService;
+
+// ==================== Singleton Pattern ====================
+
+/**
+ * Generate config key for singleton cache
+ */
+function getConfigKey(config: EmbeddingConfig): string {
+  const provider = config.provider || 'google';
+  const modelName = typeof config.model === 'string'
+    ? config.model
+    : (config.modelName || 'default');
+  const dimensions = config.dimensions || 'default';
+  return `${provider}:${modelName}:${dimensions}`;
+}
+
+/** Singleton cache */
+const sharedInstances = new Map<string, EmbeddingService>();
+
+/**
+ * Get or create a shared EmbeddingService instance (Singleton)
+ *
+ * All clients with same provider/model/dimensions share one instance.
+ * Reduces memory usage and connection overhead.
+ *
+ * @example
+ * ```typescript
+ * // Instead of: new EmbeddingService({ apiKey, modelName })
+ * const embedding = getSharedEmbeddingService({ apiKey, modelName });
+ * ```
+ */
+export function getSharedEmbeddingService(config: EmbeddingConfig): EmbeddingService {
+  const key = getConfigKey(config);
+
+  let instance = sharedInstances.get(key);
+  if (!instance) {
+    console.log(`🔧 [Singleton] Creating shared EmbeddingService: ${key}`);
+    instance = new EmbeddingService(config);
+    sharedInstances.set(key, instance);
+  }
+
+  return instance;
+}
+
+/**
+ * Clear all shared instances (for testing)
+ */
+export function clearSharedEmbeddingServices(): void {
+  sharedInstances.clear();
+}
+
+/**
+ * Get singleton stats
+ */
+export function getSharedEmbeddingStats(): {
+  instanceCount: number;
+  instances: Array<{ key: string; stats: ReturnType<EmbeddingService['getStats']> }>;
+} {
+  return {
+    instanceCount: sharedInstances.size,
+    instances: Array.from(sharedInstances.entries()).map(([key, svc]) => ({
+      key,
+      stats: svc.getStats(),
+    })),
+  };
+}