@utaba/deep-memory-embeddings-openai 0.8.1 → 0.9.0

package/dist/index.cjs

@@ -26,12 +26,14 @@ module.exports = __toCommonJS(index_exports);
 
 // src/OpenAIEmbeddingProvider.ts
 var import_deep_memory = require("@utaba/deep-memory");
+var PROVIDER_NAME = "openai";
 var OpenAIEmbeddingProvider = class {
   _baseUrl;
   _model;
   _apiKey;
   _timeoutMs;
   _maxBatchSize;
+  _reportUsage;
   _dimensions;
   constructor(config) {
     this._baseUrl = config.baseUrl.replace(/\/+$/, "");
@@ -40,34 +42,66 @@ var OpenAIEmbeddingProvider = class {
     this._timeoutMs = config.timeoutMs ?? 3e4;
     this._maxBatchSize = config.maxBatchSize ?? 64;
     this._dimensions = config.dimensions;
+    this._reportUsage = (0, import_deep_memory.createSafeSink)(config.reportUsage);
   }
   async embed(text) {
-    const response = await this._request([text]);
-    const vector = response.data[0]?.embedding;
-    if (!vector) {
-      throw new import_deep_memory.ProviderError(
-        "OpenAI embeddings API returned no data",
-        "Check that the model is loaded and the endpoint is responding correctly."
-      );
+    let totalPromptTokens = 0;
+    let totalTokens = 0;
+    let calls = 0;
+    try {
+      const response = await this._request([text]);
+      totalPromptTokens += response.usage?.prompt_tokens ?? 0;
+      totalTokens += response.usage?.total_tokens ?? 0;
+      calls++;
+      const vector = response.data[0]?.embedding;
+      if (!vector) {
+        throw new import_deep_memory.ProviderError(
+          "OpenAI embeddings API returned no data",
+          "Check that the model is loaded and the endpoint is responding correctly."
+        );
+      }
+      this._resolveDimensions(vector);
+      return vector;
+    } finally {
+      this._emitUsage("embed", totalPromptTokens, totalTokens, calls);
     }
-    this._resolveDimensions(vector);
-    return vector;
   }
   async embedBatch(texts) {
     if (texts.length === 0) return [];
     const results = [];
-    for (let i = 0; i < texts.length; i += this._maxBatchSize) {
-      const chunk = texts.slice(i, i + this._maxBatchSize);
-      const response = await this._request(chunk);
-      const sorted = response.data.sort((a, b) => a.index - b.index);
-      for (const item of sorted) {
-        results.push(item.embedding);
+    let totalPromptTokens = 0;
+    let totalTokens = 0;
+    let calls = 0;
+    try {
+      for (let i = 0; i < texts.length; i += this._maxBatchSize) {
+        const chunk = texts.slice(i, i + this._maxBatchSize);
+        const response = await this._request(chunk);
+        totalPromptTokens += response.usage?.prompt_tokens ?? 0;
+        totalTokens += response.usage?.total_tokens ?? 0;
+        calls++;
+        const sorted = response.data.sort((a, b) => a.index - b.index);
+        for (const item of sorted) {
+          results.push(item.embedding);
+        }
       }
+      if (results.length > 0 && results[0]) {
+        this._resolveDimensions(results[0]);
+      }
+      return results;
+    } finally {
+      this._emitUsage("embedBatch", totalPromptTokens, totalTokens, calls);
     }
-    if (results.length > 0 && results[0]) {
-      this._resolveDimensions(results[0]);
-    }
-    return results;
+  }
+  _emitUsage(operation, promptTokens, totalTokens, calls) {
+    if (!this._reportUsage || calls === 0) return;
+    this._reportUsage({
+      provider: PROVIDER_NAME,
+      operation,
+      unit: "tokens",
+      value: totalTokens,
+      timestamp: /* @__PURE__ */ new Date(),
+      details: { promptTokens, totalTokens, calls }
+    });
   }
   dimensions() {
     if (this._dimensions === void 0) {

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts","../src/OpenAIEmbeddingProvider.ts"],"sourcesContent":["// @utaba/deep-memory-embeddings-openai — OpenAI-compatible EmbeddingProvider\n\nexport {\n  OpenAIEmbeddingProvider,\n  type OpenAIEmbeddingProviderConfig,\n} from './OpenAIEmbeddingProvider.js';\n","// OpenAI-compatible EmbeddingProvider — works with vLLM, OpenAI, Azure, Ollama, etc.\n\nimport type { EmbeddingProvider } from '@utaba/deep-memory/providers';\nimport { ProviderError } from '@utaba/deep-memory';\n\n/** Configuration for the OpenAI-compatible embedding provider */\nexport interface OpenAIEmbeddingProviderConfig {\n  /** Base URL of the embeddings API (e.g. \"http://localhost:8010\") */\n  baseUrl: string;\n\n  /** Model identifier sent in requests (e.g. \"Qwen/Qwen3-Embedding-8B\") */\n  model: string;\n\n  /** Dimensionality of the embedding vectors. Auto-detected on first call if omitted. */\n  dimensions?: number;\n\n  /** API key for authenticated endpoints. Optional for local servers. */\n  apiKey?: string;\n\n  /** Request timeout in milliseconds. Default: 30000 */\n  timeoutMs?: number;\n\n  /** Maximum texts per batch request. Default: 64 */\n  maxBatchSize?: number;\n}\n\n/** Response shape from the OpenAI /v1/embeddings endpoint */\ninterface EmbeddingsResponse {\n  object: string;\n  data: Array<{ object: string; index: number; embedding: number[] }>;\n  model: string;\n  usage: { prompt_tokens: number; total_tokens: number };\n}\n\nexport class OpenAIEmbeddingProvider implements EmbeddingProvider {\n  private readonly _baseUrl: string;\n  private readonly _model: string;\n  private readonly _apiKey: string | undefined;\n  private readonly _timeoutMs: number;\n  private readonly _maxBatchSize: number;\n  private _dimensions: number | undefined;\n\n  constructor(config: OpenAIEmbeddingProviderConfig) {\n    this._baseUrl = config.baseUrl.replace(/\\/+$/, '');\n    this._model = config.model;\n    this._apiKey = config.apiKey;\n    this._timeoutMs = config.timeoutMs ?? 30_000;\n    this._maxBatchSize = config.maxBatchSize ?? 64;\n    this._dimensions = config.dimensions;\n  }\n\n  async embed(text: string): Promise<number[]> {\n    const response = await this._request([text]);\n    const vector = response.data[0]?.embedding;\n    if (!vector) {\n      throw new ProviderError(\n        'OpenAI embeddings API returned no data',\n        'Check that the model is loaded and the endpoint is responding correctly.',\n      );\n    }\n    this._resolveDimensions(vector);\n    return vector;\n  }\n\n  async embedBatch(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) return [];\n\n    const results: number[][] = [];\n\n    // Chunk into maxBatchSize to avoid overloading the server\n    for (let i = 0; i < texts.length; i += this._maxBatchSize) {\n      const chunk = texts.slice(i, i + this._maxBatchSize);\n      const response = await this._request(chunk);\n\n      // Sort by index to guarantee order matches input\n      const sorted = response.data.sort((a, b) => a.index - b.index);\n      for (const item of sorted) {\n        results.push(item.embedding);\n      }\n    }\n\n    if (results.length > 0 && results[0]) {\n      this._resolveDimensions(results[0]);\n    }\n\n    return results;\n  }\n\n  dimensions(): number {\n    if (this._dimensions === undefined) {\n      throw new ProviderError(\n        'Embedding dimensions not yet known. Call embed() first or provide dimensions in config.',\n        'Either set dimensions in OpenAIEmbeddingProviderConfig, or call embed() once before calling dimensions().',\n      );\n    }\n    return this._dimensions;\n  }\n\n  modelId(): string {\n    return this._model;\n  }\n\n  similarity(a: number[], b: number[]): number {\n    return cosineSimilarity(a, b);\n  }\n\n  private async _request(input: string[]): Promise<EmbeddingsResponse> {\n    const url = `${this._baseUrl}/v1/embeddings`;\n\n    const headers: Record<string, string> = {\n      'Content-Type': 'application/json',\n    };\n    if (this._apiKey) {\n      headers['Authorization'] = `Bearer ${this._apiKey}`;\n    }\n\n    const requestBody: { input: string[]; model: string; dimensions?: number } = {\n      input,\n      model: this._model,\n    };\n    if (this._dimensions !== undefined) {\n      requestBody.dimensions = this._dimensions;\n    }\n    const body = JSON.stringify(requestBody);\n\n    let response: Response;\n    try {\n      response = await fetch(url, {\n        method: 'POST',\n        headers,\n        body,\n        signal: AbortSignal.timeout(this._timeoutMs),\n      });\n    } catch (error: unknown) {\n      if (error instanceof DOMException && error.name === 'TimeoutError') {\n        throw new ProviderError(\n          `Embedding request timed out after ${this._timeoutMs}ms`,\n          `Increase timeoutMs in config, reduce batch size, or check server health at ${this._baseUrl}/health`,\n        );\n      }\n      throw new ProviderError(\n        `Failed to connect to embedding server: ${error instanceof Error ? error.message : String(error)}`,\n        `Check that the server is running at ${this._baseUrl}`,\n      );\n    }\n\n    if (!response.ok) {\n      const errorBody = await response.text().catch(() => '');\n      throw new ProviderError(\n        `Embedding API returned ${response.status}: ${errorBody}`,\n        response.status === 401\n          ? 'Check your API key in OpenAIEmbeddingProviderConfig.'\n          : `Check the embedding server logs. URL: ${url}`,\n      );\n    }\n\n    return (await response.json()) as EmbeddingsResponse;\n  }\n\n  private _resolveDimensions(vector: number[]): void {\n    if (this._dimensions === undefined) {\n      this._dimensions = vector.length;\n    }\n  }\n}\n\n/** Cosine similarity between two vectors */\nfunction cosineSimilarity(a: number[], b: number[]): number {\n  let dot = 0;\n  let normA = 0;\n  let normB = 0;\n  for (let i = 0; i < a.length; i++) {\n    const ai = a[i]!;\n    const bi = b[i]!;\n    dot += ai * bi;\n    normA += ai * ai;\n    normB += bi * bi;\n  }\n  const denom = Math.sqrt(normA) * Math.sqrt(normB);\n  return denom === 0 ? 0 : dot / denom;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACGA,yBAA8B;AA+BvB,IAAM,0BAAN,MAA2D;AAAA,EAC/C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EAER,YAAY,QAAuC;AACjD,SAAK,WAAW,OAAO,QAAQ,QAAQ,QAAQ,EAAE;AACjD,SAAK,SAAS,OAAO;AACrB,SAAK,UAAU,OAAO;AACtB,SAAK,aAAa,OAAO,aAAa;AACtC,SAAK,gBAAgB,OAAO,gBAAgB;AAC5C,SAAK,cAAc,OAAO;AAAA,EAC5B;AAAA,EAEA,MAAM,MAAM,MAAiC;AAC3C,UAAM,WAAW,MAAM,KAAK,SAAS,CAAC,IAAI,CAAC;AAC3C,UAAM,SAAS,SAAS,KAAK,CAAC,GAAG;AACjC,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,SAAK,mBAAmB,MAAM;AAC9B,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,WAAW,OAAsC;AACrD,QAAI,MAAM,WAAW,EAAG,QAAO,CAAC;AAEhC,UAAM,UAAsB,CAAC;AAG7B,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,KAAK,eAAe;AACzD,YAAM,QAAQ,MAAM,MAAM,GAAG,IAAI,KAAK,aAAa;AACnD,YAAM,WAAW,MAAM,KAAK,SAAS,KAAK;AAG1C,YAAM,SAAS,SAAS,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC7D,iBAAW,QAAQ,QAAQ;AACzB,gBAAQ,KAAK,KAAK,SAAS;AAAA,MAC7B;AAAA,IACF;AAEA,QAAI,QAAQ,SAAS,KAAK,QAAQ,CAAC,GAAG;AACpC,WAAK,mBAAmB,QAAQ,CAAC,CAAC;AAAA,IACpC;AAEA,WAAO;AAAA,EACT;AAAA,EAEA,aAAqB;AACnB,QAAI,KAAK,gBAAgB,QAAW;AAClC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,UAAkB;AAChB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,WAAW,GAAa,GAAqB;AAC3C,WAAO,iBAAiB,GAAG,CAAC;AAAA,EAC9B;AAAA,EAEA,MAAc,SAAS,OAA8C;AACnE,UAAM,MAAM,GAAG,KAAK,QAAQ;AAE5B,UAAM,UAAkC;AAAA,MACtC,gBAAgB;AAAA,IAClB;AACA,QAAI,KAAK,SAAS;AAChB,cAAQ,eAAe,IAAI,UAAU,KAAK,OAAO;AAAA,IACnD;AAEA,UAAM,cAAuE;AAAA,MAC3E;AAAA,MACA,OAAO,KAAK;AAAA,IACd;AACA,QAAI,KAAK,gBAAgB,QAAW;AAClC,kBAAY,aAAa,KAAK;AAAA,IAChC;AACA,UAAM,OAAO,KAAK,UAAU,WAAW;AAEvC,QAAI;AACJ,QAAI;AACF,iBAAW,MAAM,MAAM,KAAK;AAAA,QAC1B,QAAQ;AAAA,QACR;AAAA,QACA;AAAA,QACA,QAAQ,YAAY,QAAQ,KAAK,UAAU;AAAA,MAC7C,CAAC;AAAA,IACH,SAAS,OAAgB;AACvB,UAAI,iBAAiB,gBAAgB,MAAM,SAAS,gBAAgB;AAClE,cAAM,IAAI;AAAA,UACR,qCAAqC,KAAK,UAAU;AAAA,UACpD,8EAA8E,KAAK,QAAQ;AAAA,QAC7F;AAAA,MACF;AACA,YAAM,IAAI;AAAA,QACR,0CAA0C,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,CAAC;AAAA,QAChG,uCAAuC,KAAK,QAAQ;AAAA,MACtD;AAAA,IACF;AAEA,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,YAAY,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACtD,YAAM,IAAI;AAAA,QACR,0BAA0B,SAAS,MAAM,KAAK,SAAS;AAAA,QACvD,SAAS,WAAW,MAChB,yDACA,yCAAyC,GAAG;AAAA,MAClD;AAAA,IACF;AAEA,WAAQ,MAAM,SAAS,KAAK;AAAA,EAC9B;AAAA,EAEQ,mBAAmB,QAAwB;AACjD,QAAI,KAAK,gBAAgB,QAAW;AAClC,WAAK,cAAc,OAAO;AAAA,IAC5B;AAAA,EACF;AACF;AAGA,SAAS,iBAAiB,GAAa,GAAqB;AAC1D,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC;AACd,UAAM,KAAK,EAAE,CAAC;AACd,WAAO,KAAK;AACZ,aAAS,KAAK;AACd,aAAS,KAAK;AAAA,EAChB;AACA,QAAM,QAAQ,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,KAAK;AAChD,SAAO,UAAU,IAAI,IAAI,MAAM;AACjC;","names":[]}
+{"version":3,"sources":["../src/index.ts","../src/OpenAIEmbeddingProvider.ts"],"sourcesContent":["// @utaba/deep-memory-embeddings-openai — OpenAI-compatible EmbeddingProvider\n\nexport {\n  OpenAIEmbeddingProvider,\n  type OpenAIEmbeddingProviderConfig,\n} from './OpenAIEmbeddingProvider.js';\n","// OpenAI-compatible EmbeddingProvider — works with vLLM, OpenAI, Azure, Ollama, etc.\n\nimport type { EmbeddingProvider } from '@utaba/deep-memory/providers';\nimport type { UsageSink } from '@utaba/deep-memory/types';\nimport { ProviderError, createSafeSink } from '@utaba/deep-memory';\n\nconst PROVIDER_NAME = 'openai';\n\n/** Configuration for the OpenAI-compatible embedding provider */\nexport interface OpenAIEmbeddingProviderConfig {\n  /** Base URL of the embeddings API (e.g. \"http://localhost:8010\") */\n  baseUrl: string;\n\n  /** Model identifier sent in requests (e.g. \"Qwen/Qwen3-Embedding-8B\") */\n  model: string;\n\n  /** Dimensionality of the embedding vectors. Auto-detected on first call if omitted. */\n  dimensions?: number;\n\n  /** API key for authenticated endpoints. Optional for local servers. */\n  apiKey?: string;\n\n  /** Request timeout in milliseconds. Default: 30000 */\n  timeoutMs?: number;\n\n  /** Maximum texts per batch request. Default: 64 */\n  maxBatchSize?: number;\n\n  /**\n   * Optional usage sink. When provided, the provider emits one\n   * {@link OperationUsage} record per `embed`/`embedBatch` call reporting\n   * prompt tokens consumed. Multi-chunk `embedBatch` aggregates across\n   * chunks into a single record.\n   */\n  reportUsage?: UsageSink;\n}\n\n/** Response shape from the OpenAI /v1/embeddings endpoint */\ninterface EmbeddingsResponse {\n  object: string;\n  data: Array<{ object: string; index: number; embedding: number[] }>;\n  model: string;\n  usage: { prompt_tokens: number; total_tokens: number };\n}\n\nexport class OpenAIEmbeddingProvider implements EmbeddingProvider {\n  private readonly _baseUrl: string;\n  private readonly _model: string;\n  private readonly _apiKey: string | undefined;\n  private readonly _timeoutMs: number;\n  private readonly _maxBatchSize: number;\n  private readonly _reportUsage: UsageSink | undefined;\n  private _dimensions: number | undefined;\n\n  constructor(config: OpenAIEmbeddingProviderConfig) {\n    this._baseUrl = config.baseUrl.replace(/\\/+$/, '');\n    this._model = config.model;\n    this._apiKey = config.apiKey;\n    this._timeoutMs = config.timeoutMs ?? 30_000;\n    this._maxBatchSize = config.maxBatchSize ?? 64;\n    this._dimensions = config.dimensions;\n    this._reportUsage = createSafeSink(config.reportUsage);\n  }\n\n  async embed(text: string): Promise<number[]> {\n    let totalPromptTokens = 0;\n    let totalTokens = 0;\n    let calls = 0;\n    try {\n      const response = await this._request([text]);\n      totalPromptTokens += response.usage?.prompt_tokens ?? 0;\n      totalTokens += response.usage?.total_tokens ?? 0;\n      calls++;\n      const vector = response.data[0]?.embedding;\n      if (!vector) {\n        throw new ProviderError(\n          'OpenAI embeddings API returned no data',\n          'Check that the model is loaded and the endpoint is responding correctly.',\n        );\n      }\n      this._resolveDimensions(vector);\n      return vector;\n    } finally {\n      this._emitUsage('embed', totalPromptTokens, totalTokens, calls);\n    }\n  }\n\n  async embedBatch(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) return [];\n\n    const results: number[][] = [];\n    let totalPromptTokens = 0;\n    let totalTokens = 0;\n    let calls = 0;\n\n    try {\n      // Chunk into maxBatchSize to avoid overloading the server\n      for (let i = 0; i < texts.length; i += this._maxBatchSize) {\n        const chunk = texts.slice(i, i + this._maxBatchSize);\n        const response = await this._request(chunk);\n        totalPromptTokens += response.usage?.prompt_tokens ?? 0;\n        totalTokens += response.usage?.total_tokens ?? 0;\n        calls++;\n\n        // Sort by index to guarantee order matches input\n        const sorted = response.data.sort((a, b) => a.index - b.index);\n        for (const item of sorted) {\n          results.push(item.embedding);\n        }\n      }\n\n      if (results.length > 0 && results[0]) {\n        this._resolveDimensions(results[0]);\n      }\n\n      return results;\n    } finally {\n      this._emitUsage('embedBatch', totalPromptTokens, totalTokens, calls);\n    }\n  }\n\n  private _emitUsage(operation: string, promptTokens: number, totalTokens: number, calls: number): void {\n    if (!this._reportUsage || calls === 0) return;\n    this._reportUsage({\n      provider: PROVIDER_NAME,\n      operation,\n      unit: 'tokens',\n      value: totalTokens,\n      timestamp: new Date(),\n      details: { promptTokens, totalTokens, calls },\n    });\n  }\n\n  dimensions(): number {\n    if (this._dimensions === undefined) {\n      throw new ProviderError(\n        'Embedding dimensions not yet known. Call embed() first or provide dimensions in config.',\n        'Either set dimensions in OpenAIEmbeddingProviderConfig, or call embed() once before calling dimensions().',\n      );\n    }\n    return this._dimensions;\n  }\n\n  modelId(): string {\n    return this._model;\n  }\n\n  similarity(a: number[], b: number[]): number {\n    return cosineSimilarity(a, b);\n  }\n\n  private async _request(input: string[]): Promise<EmbeddingsResponse> {\n    const url = `${this._baseUrl}/v1/embeddings`;\n\n    const headers: Record<string, string> = {\n      'Content-Type': 'application/json',\n    };\n    if (this._apiKey) {\n      headers['Authorization'] = `Bearer ${this._apiKey}`;\n    }\n\n    const requestBody: { input: string[]; model: string; dimensions?: number } = {\n      input,\n      model: this._model,\n    };\n    if (this._dimensions !== undefined) {\n      requestBody.dimensions = this._dimensions;\n    }\n    const body = JSON.stringify(requestBody);\n\n    let response: Response;\n    try {\n      response = await fetch(url, {\n        method: 'POST',\n        headers,\n        body,\n        signal: AbortSignal.timeout(this._timeoutMs),\n      });\n    } catch (error: unknown) {\n      if (error instanceof DOMException && error.name === 'TimeoutError') {\n        throw new ProviderError(\n          `Embedding request timed out after ${this._timeoutMs}ms`,\n          `Increase timeoutMs in config, reduce batch size, or check server health at ${this._baseUrl}/health`,\n        );\n      }\n      throw new ProviderError(\n        `Failed to connect to embedding server: ${error instanceof Error ? error.message : String(error)}`,\n        `Check that the server is running at ${this._baseUrl}`,\n      );\n    }\n\n    if (!response.ok) {\n      const errorBody = await response.text().catch(() => '');\n      throw new ProviderError(\n        `Embedding API returned ${response.status}: ${errorBody}`,\n        response.status === 401\n          ? 'Check your API key in OpenAIEmbeddingProviderConfig.'\n          : `Check the embedding server logs. URL: ${url}`,\n      );\n    }\n\n    return (await response.json()) as EmbeddingsResponse;\n  }\n\n  private _resolveDimensions(vector: number[]): void {\n    if (this._dimensions === undefined) {\n      this._dimensions = vector.length;\n    }\n  }\n}\n\n/** Cosine similarity between two vectors */\nfunction cosineSimilarity(a: number[], b: number[]): number {\n  let dot = 0;\n  let normA = 0;\n  let normB = 0;\n  for (let i = 0; i < a.length; i++) {\n    const ai = a[i]!;\n    const bi = b[i]!;\n    dot += ai * bi;\n    normA += ai * ai;\n    normB += bi * bi;\n  }\n  const denom = Math.sqrt(normA) * Math.sqrt(normB);\n  return denom === 0 ? 0 : dot / denom;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACIA,yBAA8C;AAE9C,IAAM,gBAAgB;AAuCf,IAAM,0BAAN,MAA2D;AAAA,EAC/C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EAER,YAAY,QAAuC;AACjD,SAAK,WAAW,OAAO,QAAQ,QAAQ,QAAQ,EAAE;AACjD,SAAK,SAAS,OAAO;AACrB,SAAK,UAAU,OAAO;AACtB,SAAK,aAAa,OAAO,aAAa;AACtC,SAAK,gBAAgB,OAAO,gBAAgB;AAC5C,SAAK,cAAc,OAAO;AAC1B,SAAK,mBAAe,mCAAe,OAAO,WAAW;AAAA,EACvD;AAAA,EAEA,MAAM,MAAM,MAAiC;AAC3C,QAAI,oBAAoB;AACxB,QAAI,cAAc;AAClB,QAAI,QAAQ;AACZ,QAAI;AACF,YAAM,WAAW,MAAM,KAAK,SAAS,CAAC,IAAI,CAAC;AAC3C,2BAAqB,SAAS,OAAO,iBAAiB;AACtD,qBAAe,SAAS,OAAO,gBAAgB;AAC/C;AACA,YAAM,SAAS,SAAS,KAAK,CAAC,GAAG;AACjC,UAAI,CAAC,QAAQ;AACX,cAAM,IAAI;AAAA,UACR;AAAA,UACA;AAAA,QACF;AAAA,MACF;AACA,WAAK,mBAAmB,MAAM;AAC9B,aAAO;AAAA,IACT,UAAE;AACA,WAAK,WAAW,SAAS,mBAAmB,aAAa,KAAK;AAAA,IAChE;AAAA,EACF;AAAA,EAEA,MAAM,WAAW,OAAsC;AACrD,QAAI,MAAM,WAAW,EAAG,QAAO,CAAC;AAEhC,UAAM,UAAsB,CAAC;AAC7B,QAAI,oBAAoB;AACxB,QAAI,cAAc;AAClB,QAAI,QAAQ;AAEZ,QAAI;AAEF,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,KAAK,eAAe;AACzD,cAAM,QAAQ,MAAM,MAAM,GAAG,IAAI,KAAK,aAAa;AACnD,cAAM,WAAW,MAAM,KAAK,SAAS,KAAK;AAC1C,6BAAqB,SAAS,OAAO,iBAAiB;AACtD,uBAAe,SAAS,OAAO,gBAAgB;AAC/C;AAGA,cAAM,SAAS,SAAS,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC7D,mBAAW,QAAQ,QAAQ;AACzB,kBAAQ,KAAK,KAAK,SAAS;AAAA,QAC7B;AAAA,MACF;AAEA,UAAI,QAAQ,SAAS,KAAK,QAAQ,CAAC,GAAG;AACpC,aAAK,mBAAmB,QAAQ,CAAC,CAAC;AAAA,MACpC;AAEA,aAAO;AAAA,IACT,UAAE;AACA,WAAK,WAAW,cAAc,mBAAmB,aAAa,KAAK;AAAA,IACrE;AAAA,EACF;AAAA,EAEQ,WAAW,WAAmB,cAAsB,aAAqB,OAAqB;AACpG,QAAI,CAAC,KAAK,gBAAgB,UAAU,EAAG;AACvC,SAAK,aAAa;AAAA,MAChB,UAAU;AAAA,MACV;AAAA,MACA,MAAM;AAAA,MACN,OAAO;AAAA,MACP,WAAW,oBAAI,KAAK;AAAA,MACpB,SAAS,EAAE,cAAc,aAAa,MAAM;AAAA,IAC9C,CAAC;AAAA,EACH;AAAA,EAEA,aAAqB;AACnB,QAAI,KAAK,gBAAgB,QAAW;AAClC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,UAAkB;AAChB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,WAAW,GAAa,GAAqB;AAC3C,WAAO,iBAAiB,GAAG,CAAC;AAAA,EAC9B;AAAA,EAEA,MAAc,SAAS,OAA8C;AACnE,UAAM,MAAM,GAAG,KAAK,QAAQ;AAE5B,UAAM,UAAkC;AAAA,MACtC,gBAAgB;AAAA,IAClB;AACA,QAAI,KAAK,SAAS;AAChB,cAAQ,eAAe,IAAI,UAAU,KAAK,OAAO;AAAA,IACnD;AAEA,UAAM,cAAuE;AAAA,MAC3E;AAAA,MACA,OAAO,KAAK;AAAA,IACd;AACA,QAAI,KAAK,gBAAgB,QAAW;AAClC,kBAAY,aAAa,KAAK;AAAA,IAChC;AACA,UAAM,OAAO,KAAK,UAAU,WAAW;AAEvC,QAAI;AACJ,QAAI;AACF,iBAAW,MAAM,MAAM,KAAK;AAAA,QAC1B,QAAQ;AAAA,QACR;AAAA,QACA;AAAA,QACA,QAAQ,YAAY,QAAQ,KAAK,UAAU;AAAA,MAC7C,CAAC;AAAA,IACH,SAAS,OAAgB;AACvB,UAAI,iBAAiB,gBAAgB,MAAM,SAAS,gBAAgB;AAClE,cAAM,IAAI;AAAA,UACR,qCAAqC,KAAK,UAAU;AAAA,UACpD,8EAA8E,KAAK,QAAQ;AAAA,QAC7F;AAAA,MACF;AACA,YAAM,IAAI;AAAA,QACR,0CAA0C,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,CAAC;AAAA,QAChG,uCAAuC,KAAK,QAAQ;AAAA,MACtD;AAAA,IACF;AAEA,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,YAAY,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACtD,YAAM,IAAI;AAAA,QACR,0BAA0B,SAAS,MAAM,KAAK,SAAS;AAAA,QACvD,SAAS,WAAW,MAChB,yDACA,yCAAyC,GAAG;AAAA,MAClD;AAAA,IACF;AAEA,WAAQ,MAAM,SAAS,KAAK;AAAA,EAC9B;AAAA,EAEQ,mBAAmB,QAAwB;AACjD,QAAI,KAAK,gBAAgB,QAAW;AAClC,WAAK,cAAc,OAAO;AAAA,IAC5B;AAAA,EACF;AACF;AAGA,SAAS,iBAAiB,GAAa,GAAqB;AAC1D,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC;AACd,UAAM,KAAK,EAAE,CAAC;AACd,WAAO,KAAK;AACZ,aAAS,KAAK;AACd,aAAS,KAAK;AAAA,EAChB;AACA,QAAM,QAAQ,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,KAAK;AAChD,SAAO,UAAU,IAAI,IAAI,MAAM;AACjC;","names":[]}

package/dist/index.d.cts

@@ -1,4 +1,5 @@
 import { EmbeddingProvider } from '@utaba/deep-memory/providers';
+import { UsageSink } from '@utaba/deep-memory/types';
 
 /** Configuration for the OpenAI-compatible embedding provider */
 interface OpenAIEmbeddingProviderConfig {
@@ -14,6 +15,13 @@ interface OpenAIEmbeddingProviderConfig {
     timeoutMs?: number;
     /** Maximum texts per batch request. Default: 64 */
     maxBatchSize?: number;
+    /**
+     * Optional usage sink. When provided, the provider emits one
+     * {@link OperationUsage} record per `embed`/`embedBatch` call reporting
+     * prompt tokens consumed. Multi-chunk `embedBatch` aggregates across
+     * chunks into a single record.
+     */
+    reportUsage?: UsageSink;
 }
 declare class OpenAIEmbeddingProvider implements EmbeddingProvider {
     private readonly _baseUrl;
@@ -21,10 +29,12 @@ declare class OpenAIEmbeddingProvider implements EmbeddingProvider {
     private readonly _apiKey;
     private readonly _timeoutMs;
     private readonly _maxBatchSize;
+    private readonly _reportUsage;
     private _dimensions;
     constructor(config: OpenAIEmbeddingProviderConfig);
     embed(text: string): Promise<number[]>;
     embedBatch(texts: string[]): Promise<number[][]>;
+    private _emitUsage;
     dimensions(): number;
     modelId(): string;
     similarity(a: number[], b: number[]): number;

package/dist/index.d.ts

@@ -1,4 +1,5 @@
 import { EmbeddingProvider } from '@utaba/deep-memory/providers';
+import { UsageSink } from '@utaba/deep-memory/types';
 
 /** Configuration for the OpenAI-compatible embedding provider */
 interface OpenAIEmbeddingProviderConfig {
@@ -14,6 +15,13 @@ interface OpenAIEmbeddingProviderConfig {
     timeoutMs?: number;
     /** Maximum texts per batch request. Default: 64 */
     maxBatchSize?: number;
+    /**
+     * Optional usage sink. When provided, the provider emits one
+     * {@link OperationUsage} record per `embed`/`embedBatch` call reporting
+     * prompt tokens consumed. Multi-chunk `embedBatch` aggregates across
+     * chunks into a single record.
+     */
+    reportUsage?: UsageSink;
 }
 declare class OpenAIEmbeddingProvider implements EmbeddingProvider {
     private readonly _baseUrl;
@@ -21,10 +29,12 @@ declare class OpenAIEmbeddingProvider implements EmbeddingProvider {
     private readonly _apiKey;
     private readonly _timeoutMs;
     private readonly _maxBatchSize;
+    private readonly _reportUsage;
     private _dimensions;
     constructor(config: OpenAIEmbeddingProviderConfig);
     embed(text: string): Promise<number[]>;
     embedBatch(texts: string[]): Promise<number[][]>;
+    private _emitUsage;
     dimensions(): number;
     modelId(): string;
     similarity(a: number[], b: number[]): number;

package/dist/index.js

@@ -1,11 +1,13 @@
 // src/OpenAIEmbeddingProvider.ts
-import { ProviderError } from "@utaba/deep-memory";
+import { ProviderError, createSafeSink } from "@utaba/deep-memory";
+var PROVIDER_NAME = "openai";
 var OpenAIEmbeddingProvider = class {
   _baseUrl;
   _model;
   _apiKey;
   _timeoutMs;
   _maxBatchSize;
+  _reportUsage;
   _dimensions;
   constructor(config) {
     this._baseUrl = config.baseUrl.replace(/\/+$/, "");
@@ -14,34 +16,66 @@ var OpenAIEmbeddingProvider = class {
     this._timeoutMs = config.timeoutMs ?? 3e4;
     this._maxBatchSize = config.maxBatchSize ?? 64;
     this._dimensions = config.dimensions;
+    this._reportUsage = createSafeSink(config.reportUsage);
   }
   async embed(text) {
-    const response = await this._request([text]);
-    const vector = response.data[0]?.embedding;
-    if (!vector) {
-      throw new ProviderError(
-        "OpenAI embeddings API returned no data",
-        "Check that the model is loaded and the endpoint is responding correctly."
-      );
+    let totalPromptTokens = 0;
+    let totalTokens = 0;
+    let calls = 0;
+    try {
+      const response = await this._request([text]);
+      totalPromptTokens += response.usage?.prompt_tokens ?? 0;
+      totalTokens += response.usage?.total_tokens ?? 0;
+      calls++;
+      const vector = response.data[0]?.embedding;
+      if (!vector) {
+        throw new ProviderError(
+          "OpenAI embeddings API returned no data",
+          "Check that the model is loaded and the endpoint is responding correctly."
+        );
+      }
+      this._resolveDimensions(vector);
+      return vector;
+    } finally {
+      this._emitUsage("embed", totalPromptTokens, totalTokens, calls);
     }
-    this._resolveDimensions(vector);
-    return vector;
   }
   async embedBatch(texts) {
     if (texts.length === 0) return [];
     const results = [];
-    for (let i = 0; i < texts.length; i += this._maxBatchSize) {
-      const chunk = texts.slice(i, i + this._maxBatchSize);
-      const response = await this._request(chunk);
-      const sorted = response.data.sort((a, b) => a.index - b.index);
-      for (const item of sorted) {
-        results.push(item.embedding);
+    let totalPromptTokens = 0;
+    let totalTokens = 0;
+    let calls = 0;
+    try {
+      for (let i = 0; i < texts.length; i += this._maxBatchSize) {
+        const chunk = texts.slice(i, i + this._maxBatchSize);
+        const response = await this._request(chunk);
+        totalPromptTokens += response.usage?.prompt_tokens ?? 0;
+        totalTokens += response.usage?.total_tokens ?? 0;
+        calls++;
+        const sorted = response.data.sort((a, b) => a.index - b.index);
+        for (const item of sorted) {
+          results.push(item.embedding);
+        }
       }
+      if (results.length > 0 && results[0]) {
+        this._resolveDimensions(results[0]);
+      }
+      return results;
+    } finally {
+      this._emitUsage("embedBatch", totalPromptTokens, totalTokens, calls);
     }
-    if (results.length > 0 && results[0]) {
-      this._resolveDimensions(results[0]);
-    }
-    return results;
+  }
+  _emitUsage(operation, promptTokens, totalTokens, calls) {
+    if (!this._reportUsage || calls === 0) return;
+    this._reportUsage({
+      provider: PROVIDER_NAME,
+      operation,
+      unit: "tokens",
+      value: totalTokens,
+      timestamp: /* @__PURE__ */ new Date(),
+      details: { promptTokens, totalTokens, calls }
+    });
   }
   dimensions() {
     if (this._dimensions === void 0) {

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/OpenAIEmbeddingProvider.ts"],"sourcesContent":["// OpenAI-compatible EmbeddingProvider — works with vLLM, OpenAI, Azure, Ollama, etc.\n\nimport type { EmbeddingProvider } from '@utaba/deep-memory/providers';\nimport { ProviderError } from '@utaba/deep-memory';\n\n/** Configuration for the OpenAI-compatible embedding provider */\nexport interface OpenAIEmbeddingProviderConfig {\n  /** Base URL of the embeddings API (e.g. \"http://localhost:8010\") */\n  baseUrl: string;\n\n  /** Model identifier sent in requests (e.g. \"Qwen/Qwen3-Embedding-8B\") */\n  model: string;\n\n  /** Dimensionality of the embedding vectors. Auto-detected on first call if omitted. */\n  dimensions?: number;\n\n  /** API key for authenticated endpoints. Optional for local servers. */\n  apiKey?: string;\n\n  /** Request timeout in milliseconds. Default: 30000 */\n  timeoutMs?: number;\n\n  /** Maximum texts per batch request. Default: 64 */\n  maxBatchSize?: number;\n}\n\n/** Response shape from the OpenAI /v1/embeddings endpoint */\ninterface EmbeddingsResponse {\n  object: string;\n  data: Array<{ object: string; index: number; embedding: number[] }>;\n  model: string;\n  usage: { prompt_tokens: number; total_tokens: number };\n}\n\nexport class OpenAIEmbeddingProvider implements EmbeddingProvider {\n  private readonly _baseUrl: string;\n  private readonly _model: string;\n  private readonly _apiKey: string | undefined;\n  private readonly _timeoutMs: number;\n  private readonly _maxBatchSize: number;\n  private _dimensions: number | undefined;\n\n  constructor(config: OpenAIEmbeddingProviderConfig) {\n    this._baseUrl = config.baseUrl.replace(/\\/+$/, '');\n    this._model = config.model;\n    this._apiKey = config.apiKey;\n    this._timeoutMs = config.timeoutMs ?? 30_000;\n    this._maxBatchSize = config.maxBatchSize ?? 64;\n    this._dimensions = config.dimensions;\n  }\n\n  async embed(text: string): Promise<number[]> {\n    const response = await this._request([text]);\n    const vector = response.data[0]?.embedding;\n    if (!vector) {\n      throw new ProviderError(\n        'OpenAI embeddings API returned no data',\n        'Check that the model is loaded and the endpoint is responding correctly.',\n      );\n    }\n    this._resolveDimensions(vector);\n    return vector;\n  }\n\n  async embedBatch(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) return [];\n\n    const results: number[][] = [];\n\n    // Chunk into maxBatchSize to avoid overloading the server\n    for (let i = 0; i < texts.length; i += this._maxBatchSize) {\n      const chunk = texts.slice(i, i + this._maxBatchSize);\n      const response = await this._request(chunk);\n\n      // Sort by index to guarantee order matches input\n      const sorted = response.data.sort((a, b) => a.index - b.index);\n      for (const item of sorted) {\n        results.push(item.embedding);\n      }\n    }\n\n    if (results.length > 0 && results[0]) {\n      this._resolveDimensions(results[0]);\n    }\n\n    return results;\n  }\n\n  dimensions(): number {\n    if (this._dimensions === undefined) {\n      throw new ProviderError(\n        'Embedding dimensions not yet known. Call embed() first or provide dimensions in config.',\n        'Either set dimensions in OpenAIEmbeddingProviderConfig, or call embed() once before calling dimensions().',\n      );\n    }\n    return this._dimensions;\n  }\n\n  modelId(): string {\n    return this._model;\n  }\n\n  similarity(a: number[], b: number[]): number {\n    return cosineSimilarity(a, b);\n  }\n\n  private async _request(input: string[]): Promise<EmbeddingsResponse> {\n    const url = `${this._baseUrl}/v1/embeddings`;\n\n    const headers: Record<string, string> = {\n      'Content-Type': 'application/json',\n    };\n    if (this._apiKey) {\n      headers['Authorization'] = `Bearer ${this._apiKey}`;\n    }\n\n    const requestBody: { input: string[]; model: string; dimensions?: number } = {\n      input,\n      model: this._model,\n    };\n    if (this._dimensions !== undefined) {\n      requestBody.dimensions = this._dimensions;\n    }\n    const body = JSON.stringify(requestBody);\n\n    let response: Response;\n    try {\n      response = await fetch(url, {\n        method: 'POST',\n        headers,\n        body,\n        signal: AbortSignal.timeout(this._timeoutMs),\n      });\n    } catch (error: unknown) {\n      if (error instanceof DOMException && error.name === 'TimeoutError') {\n        throw new ProviderError(\n          `Embedding request timed out after ${this._timeoutMs}ms`,\n          `Increase timeoutMs in config, reduce batch size, or check server health at ${this._baseUrl}/health`,\n        );\n      }\n      throw new ProviderError(\n        `Failed to connect to embedding server: ${error instanceof Error ? error.message : String(error)}`,\n        `Check that the server is running at ${this._baseUrl}`,\n      );\n    }\n\n    if (!response.ok) {\n      const errorBody = await response.text().catch(() => '');\n      throw new ProviderError(\n        `Embedding API returned ${response.status}: ${errorBody}`,\n        response.status === 401\n          ? 'Check your API key in OpenAIEmbeddingProviderConfig.'\n          : `Check the embedding server logs. URL: ${url}`,\n      );\n    }\n\n    return (await response.json()) as EmbeddingsResponse;\n  }\n\n  private _resolveDimensions(vector: number[]): void {\n    if (this._dimensions === undefined) {\n      this._dimensions = vector.length;\n    }\n  }\n}\n\n/** Cosine similarity between two vectors */\nfunction cosineSimilarity(a: number[], b: number[]): number {\n  let dot = 0;\n  let normA = 0;\n  let normB = 0;\n  for (let i = 0; i < a.length; i++) {\n    const ai = a[i]!;\n    const bi = b[i]!;\n    dot += ai * bi;\n    normA += ai * ai;\n    normB += bi * bi;\n  }\n  const denom = Math.sqrt(normA) * Math.sqrt(normB);\n  return denom === 0 ? 0 : dot / denom;\n}\n"],"mappings":";AAGA,SAAS,qBAAqB;AA+BvB,IAAM,0BAAN,MAA2D;AAAA,EAC/C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EAER,YAAY,QAAuC;AACjD,SAAK,WAAW,OAAO,QAAQ,QAAQ,QAAQ,EAAE;AACjD,SAAK,SAAS,OAAO;AACrB,SAAK,UAAU,OAAO;AACtB,SAAK,aAAa,OAAO,aAAa;AACtC,SAAK,gBAAgB,OAAO,gBAAgB;AAC5C,SAAK,cAAc,OAAO;AAAA,EAC5B;AAAA,EAEA,MAAM,MAAM,MAAiC;AAC3C,UAAM,WAAW,MAAM,KAAK,SAAS,CAAC,IAAI,CAAC;AAC3C,UAAM,SAAS,SAAS,KAAK,CAAC,GAAG;AACjC,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,SAAK,mBAAmB,MAAM;AAC9B,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,WAAW,OAAsC;AACrD,QAAI,MAAM,WAAW,EAAG,QAAO,CAAC;AAEhC,UAAM,UAAsB,CAAC;AAG7B,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,KAAK,eAAe;AACzD,YAAM,QAAQ,MAAM,MAAM,GAAG,IAAI,KAAK,aAAa;AACnD,YAAM,WAAW,MAAM,KAAK,SAAS,KAAK;AAG1C,YAAM,SAAS,SAAS,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC7D,iBAAW,QAAQ,QAAQ;AACzB,gBAAQ,KAAK,KAAK,SAAS;AAAA,MAC7B;AAAA,IACF;AAEA,QAAI,QAAQ,SAAS,KAAK,QAAQ,CAAC,GAAG;AACpC,WAAK,mBAAmB,QAAQ,CAAC,CAAC;AAAA,IACpC;AAEA,WAAO;AAAA,EACT;AAAA,EAEA,aAAqB;AACnB,QAAI,KAAK,gBAAgB,QAAW;AAClC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,UAAkB;AAChB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,WAAW,GAAa,GAAqB;AAC3C,WAAO,iBAAiB,GAAG,CAAC;AAAA,EAC9B;AAAA,EAEA,MAAc,SAAS,OAA8C;AACnE,UAAM,MAAM,GAAG,KAAK,QAAQ;AAE5B,UAAM,UAAkC;AAAA,MACtC,gBAAgB;AAAA,IAClB;AACA,QAAI,KAAK,SAAS;AAChB,cAAQ,eAAe,IAAI,UAAU,KAAK,OAAO;AAAA,IACnD;AAEA,UAAM,cAAuE;AAAA,MAC3E;AAAA,MACA,OAAO,KAAK;AAAA,IACd;AACA,QAAI,KAAK,gBAAgB,QAAW;AAClC,kBAAY,aAAa,KAAK;AAAA,IAChC;AACA,UAAM,OAAO,KAAK,UAAU,WAAW;AAEvC,QAAI;AACJ,QAAI;AACF,iBAAW,MAAM,MAAM,KAAK;AAAA,QAC1B,QAAQ;AAAA,QACR;AAAA,QACA;AAAA,QACA,QAAQ,YAAY,QAAQ,KAAK,UAAU;AAAA,MAC7C,CAAC;AAAA,IACH,SAAS,OAAgB;AACvB,UAAI,iBAAiB,gBAAgB,MAAM,SAAS,gBAAgB;AAClE,cAAM,IAAI;AAAA,UACR,qCAAqC,KAAK,UAAU;AAAA,UACpD,8EAA8E,KAAK,QAAQ;AAAA,QAC7F;AAAA,MACF;AACA,YAAM,IAAI;AAAA,QACR,0CAA0C,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,CAAC;AAAA,QAChG,uCAAuC,KAAK,QAAQ;AAAA,MACtD;AAAA,IACF;AAEA,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,YAAY,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACtD,YAAM,IAAI;AAAA,QACR,0BAA0B,SAAS,MAAM,KAAK,SAAS;AAAA,QACvD,SAAS,WAAW,MAChB,yDACA,yCAAyC,GAAG;AAAA,MAClD;AAAA,IACF;AAEA,WAAQ,MAAM,SAAS,KAAK;AAAA,EAC9B;AAAA,EAEQ,mBAAmB,QAAwB;AACjD,QAAI,KAAK,gBAAgB,QAAW;AAClC,WAAK,cAAc,OAAO;AAAA,IAC5B;AAAA,EACF;AACF;AAGA,SAAS,iBAAiB,GAAa,GAAqB;AAC1D,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC;AACd,UAAM,KAAK,EAAE,CAAC;AACd,WAAO,KAAK;AACZ,aAAS,KAAK;AACd,aAAS,KAAK;AAAA,EAChB;AACA,QAAM,QAAQ,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,KAAK;AAChD,SAAO,UAAU,IAAI,IAAI,MAAM;AACjC;","names":[]}
+{"version":3,"sources":["../src/OpenAIEmbeddingProvider.ts"],"sourcesContent":["// OpenAI-compatible EmbeddingProvider — works with vLLM, OpenAI, Azure, Ollama, etc.\n\nimport type { EmbeddingProvider } from '@utaba/deep-memory/providers';\nimport type { UsageSink } from '@utaba/deep-memory/types';\nimport { ProviderError, createSafeSink } from '@utaba/deep-memory';\n\nconst PROVIDER_NAME = 'openai';\n\n/** Configuration for the OpenAI-compatible embedding provider */\nexport interface OpenAIEmbeddingProviderConfig {\n  /** Base URL of the embeddings API (e.g. \"http://localhost:8010\") */\n  baseUrl: string;\n\n  /** Model identifier sent in requests (e.g. \"Qwen/Qwen3-Embedding-8B\") */\n  model: string;\n\n  /** Dimensionality of the embedding vectors. Auto-detected on first call if omitted. */\n  dimensions?: number;\n\n  /** API key for authenticated endpoints. Optional for local servers. */\n  apiKey?: string;\n\n  /** Request timeout in milliseconds. Default: 30000 */\n  timeoutMs?: number;\n\n  /** Maximum texts per batch request. Default: 64 */\n  maxBatchSize?: number;\n\n  /**\n   * Optional usage sink. When provided, the provider emits one\n   * {@link OperationUsage} record per `embed`/`embedBatch` call reporting\n   * prompt tokens consumed. Multi-chunk `embedBatch` aggregates across\n   * chunks into a single record.\n   */\n  reportUsage?: UsageSink;\n}\n\n/** Response shape from the OpenAI /v1/embeddings endpoint */\ninterface EmbeddingsResponse {\n  object: string;\n  data: Array<{ object: string; index: number; embedding: number[] }>;\n  model: string;\n  usage: { prompt_tokens: number; total_tokens: number };\n}\n\nexport class OpenAIEmbeddingProvider implements EmbeddingProvider {\n  private readonly _baseUrl: string;\n  private readonly _model: string;\n  private readonly _apiKey: string | undefined;\n  private readonly _timeoutMs: number;\n  private readonly _maxBatchSize: number;\n  private readonly _reportUsage: UsageSink | undefined;\n  private _dimensions: number | undefined;\n\n  constructor(config: OpenAIEmbeddingProviderConfig) {\n    this._baseUrl = config.baseUrl.replace(/\\/+$/, '');\n    this._model = config.model;\n    this._apiKey = config.apiKey;\n    this._timeoutMs = config.timeoutMs ?? 30_000;\n    this._maxBatchSize = config.maxBatchSize ?? 64;\n    this._dimensions = config.dimensions;\n    this._reportUsage = createSafeSink(config.reportUsage);\n  }\n\n  async embed(text: string): Promise<number[]> {\n    let totalPromptTokens = 0;\n    let totalTokens = 0;\n    let calls = 0;\n    try {\n      const response = await this._request([text]);\n      totalPromptTokens += response.usage?.prompt_tokens ?? 0;\n      totalTokens += response.usage?.total_tokens ?? 0;\n      calls++;\n      const vector = response.data[0]?.embedding;\n      if (!vector) {\n        throw new ProviderError(\n          'OpenAI embeddings API returned no data',\n          'Check that the model is loaded and the endpoint is responding correctly.',\n        );\n      }\n      this._resolveDimensions(vector);\n      return vector;\n    } finally {\n      this._emitUsage('embed', totalPromptTokens, totalTokens, calls);\n    }\n  }\n\n  async embedBatch(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) return [];\n\n    const results: number[][] = [];\n    let totalPromptTokens = 0;\n    let totalTokens = 0;\n    let calls = 0;\n\n    try {\n      // Chunk into maxBatchSize to avoid overloading the server\n      for (let i = 0; i < texts.length; i += this._maxBatchSize) {\n        const chunk = texts.slice(i, i + this._maxBatchSize);\n        const response = await this._request(chunk);\n        totalPromptTokens += response.usage?.prompt_tokens ?? 0;\n        totalTokens += response.usage?.total_tokens ?? 0;\n        calls++;\n\n        // Sort by index to guarantee order matches input\n        const sorted = response.data.sort((a, b) => a.index - b.index);\n        for (const item of sorted) {\n          results.push(item.embedding);\n        }\n      }\n\n      if (results.length > 0 && results[0]) {\n        this._resolveDimensions(results[0]);\n      }\n\n      return results;\n    } finally {\n      this._emitUsage('embedBatch', totalPromptTokens, totalTokens, calls);\n    }\n  }\n\n  private _emitUsage(operation: string, promptTokens: number, totalTokens: number, calls: number): void {\n    if (!this._reportUsage || calls === 0) return;\n    this._reportUsage({\n      provider: PROVIDER_NAME,\n      operation,\n      unit: 'tokens',\n      value: totalTokens,\n      timestamp: new Date(),\n      details: { promptTokens, totalTokens, calls },\n    });\n  }\n\n  dimensions(): number {\n    if (this._dimensions === undefined) {\n      throw new ProviderError(\n        'Embedding dimensions not yet known. Call embed() first or provide dimensions in config.',\n        'Either set dimensions in OpenAIEmbeddingProviderConfig, or call embed() once before calling dimensions().',\n      );\n    }\n    return this._dimensions;\n  }\n\n  modelId(): string {\n    return this._model;\n  }\n\n  similarity(a: number[], b: number[]): number {\n    return cosineSimilarity(a, b);\n  }\n\n  private async _request(input: string[]): Promise<EmbeddingsResponse> {\n    const url = `${this._baseUrl}/v1/embeddings`;\n\n    const headers: Record<string, string> = {\n      'Content-Type': 'application/json',\n    };\n    if (this._apiKey) {\n      headers['Authorization'] = `Bearer ${this._apiKey}`;\n    }\n\n    const requestBody: { input: string[]; model: string; dimensions?: number } = {\n      input,\n      model: this._model,\n    };\n    if (this._dimensions !== undefined) {\n      requestBody.dimensions = this._dimensions;\n    }\n    const body = JSON.stringify(requestBody);\n\n    let response: Response;\n    try {\n      response = await fetch(url, {\n        method: 'POST',\n        headers,\n        body,\n        signal: AbortSignal.timeout(this._timeoutMs),\n      });\n    } catch (error: unknown) {\n      if (error instanceof DOMException && error.name === 'TimeoutError') {\n        throw new ProviderError(\n          `Embedding request timed out after ${this._timeoutMs}ms`,\n          `Increase timeoutMs in config, reduce batch size, or check server health at ${this._baseUrl}/health`,\n        );\n      }\n      throw new ProviderError(\n        `Failed to connect to embedding server: ${error instanceof Error ? error.message : String(error)}`,\n        `Check that the server is running at ${this._baseUrl}`,\n      );\n    }\n\n    if (!response.ok) {\n      const errorBody = await response.text().catch(() => '');\n      throw new ProviderError(\n        `Embedding API returned ${response.status}: ${errorBody}`,\n        response.status === 401\n          ? 'Check your API key in OpenAIEmbeddingProviderConfig.'\n          : `Check the embedding server logs. URL: ${url}`,\n      );\n    }\n\n    return (await response.json()) as EmbeddingsResponse;\n  }\n\n  private _resolveDimensions(vector: number[]): void {\n    if (this._dimensions === undefined) {\n      this._dimensions = vector.length;\n    }\n  }\n}\n\n/** Cosine similarity between two vectors */\nfunction cosineSimilarity(a: number[], b: number[]): number {\n  let dot = 0;\n  let normA = 0;\n  let normB = 0;\n  for (let i = 0; i < a.length; i++) {\n    const ai = a[i]!;\n    const bi = b[i]!;\n    dot += ai * bi;\n    normA += ai * ai;\n    normB += bi * bi;\n  }\n  const denom = Math.sqrt(normA) * Math.sqrt(normB);\n  return denom === 0 ? 0 : dot / denom;\n}\n"],"mappings":";AAIA,SAAS,eAAe,sBAAsB;AAE9C,IAAM,gBAAgB;AAuCf,IAAM,0BAAN,MAA2D;AAAA,EAC/C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EAER,YAAY,QAAuC;AACjD,SAAK,WAAW,OAAO,QAAQ,QAAQ,QAAQ,EAAE;AACjD,SAAK,SAAS,OAAO;AACrB,SAAK,UAAU,OAAO;AACtB,SAAK,aAAa,OAAO,aAAa;AACtC,SAAK,gBAAgB,OAAO,gBAAgB;AAC5C,SAAK,cAAc,OAAO;AAC1B,SAAK,eAAe,eAAe,OAAO,WAAW;AAAA,EACvD;AAAA,EAEA,MAAM,MAAM,MAAiC;AAC3C,QAAI,oBAAoB;AACxB,QAAI,cAAc;AAClB,QAAI,QAAQ;AACZ,QAAI;AACF,YAAM,WAAW,MAAM,KAAK,SAAS,CAAC,IAAI,CAAC;AAC3C,2BAAqB,SAAS,OAAO,iBAAiB;AACtD,qBAAe,SAAS,OAAO,gBAAgB;AAC/C;AACA,YAAM,SAAS,SAAS,KAAK,CAAC,GAAG;AACjC,UAAI,CAAC,QAAQ;AACX,cAAM,IAAI;AAAA,UACR;AAAA,UACA;AAAA,QACF;AAAA,MACF;AACA,WAAK,mBAAmB,MAAM;AAC9B,aAAO;AAAA,IACT,UAAE;AACA,WAAK,WAAW,SAAS,mBAAmB,aAAa,KAAK;AAAA,IAChE;AAAA,EACF;AAAA,EAEA,MAAM,WAAW,OAAsC;AACrD,QAAI,MAAM,WAAW,EAAG,QAAO,CAAC;AAEhC,UAAM,UAAsB,CAAC;AAC7B,QAAI,oBAAoB;AACxB,QAAI,cAAc;AAClB,QAAI,QAAQ;AAEZ,QAAI;AAEF,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,KAAK,eAAe;AACzD,cAAM,QAAQ,MAAM,MAAM,GAAG,IAAI,KAAK,aAAa;AACnD,cAAM,WAAW,MAAM,KAAK,SAAS,KAAK;AAC1C,6BAAqB,SAAS,OAAO,iBAAiB;AACtD,uBAAe,SAAS,OAAO,gBAAgB;AAC/C;AAGA,cAAM,SAAS,SAAS,KAAK,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC7D,mBAAW,QAAQ,QAAQ;AACzB,kBAAQ,KAAK,KAAK,SAAS;AAAA,QAC7B;AAAA,MACF;AAEA,UAAI,QAAQ,SAAS,KAAK,QAAQ,CAAC,GAAG;AACpC,aAAK,mBAAmB,QAAQ,CAAC,CAAC;AAAA,MACpC;AAEA,aAAO;AAAA,IACT,UAAE;AACA,WAAK,WAAW,cAAc,mBAAmB,aAAa,KAAK;AAAA,IACrE;AAAA,EACF;AAAA,EAEQ,WAAW,WAAmB,cAAsB,aAAqB,OAAqB;AACpG,QAAI,CAAC,KAAK,gBAAgB,UAAU,EAAG;AACvC,SAAK,aAAa;AAAA,MAChB,UAAU;AAAA,MACV;AAAA,MACA,MAAM;AAAA,MACN,OAAO;AAAA,MACP,WAAW,oBAAI,KAAK;AAAA,MACpB,SAAS,EAAE,cAAc,aAAa,MAAM;AAAA,IAC9C,CAAC;AAAA,EACH;AAAA,EAEA,aAAqB;AACnB,QAAI,KAAK,gBAAgB,QAAW;AAClC,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,UAAkB;AAChB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,WAAW,GAAa,GAAqB;AAC3C,WAAO,iBAAiB,GAAG,CAAC;AAAA,EAC9B;AAAA,EAEA,MAAc,SAAS,OAA8C;AACnE,UAAM,MAAM,GAAG,KAAK,QAAQ;AAE5B,UAAM,UAAkC;AAAA,MACtC,gBAAgB;AAAA,IAClB;AACA,QAAI,KAAK,SAAS;AAChB,cAAQ,eAAe,IAAI,UAAU,KAAK,OAAO;AAAA,IACnD;AAEA,UAAM,cAAuE;AAAA,MAC3E;AAAA,MACA,OAAO,KAAK;AAAA,IACd;AACA,QAAI,KAAK,gBAAgB,QAAW;AAClC,kBAAY,aAAa,KAAK;AAAA,IAChC;AACA,UAAM,OAAO,KAAK,UAAU,WAAW;AAEvC,QAAI;AACJ,QAAI;AACF,iBAAW,MAAM,MAAM,KAAK;AAAA,QAC1B,QAAQ;AAAA,QACR;AAAA,QACA;AAAA,QACA,QAAQ,YAAY,QAAQ,KAAK,UAAU;AAAA,MAC7C,CAAC;AAAA,IACH,SAAS,OAAgB;AACvB,UAAI,iBAAiB,gBAAgB,MAAM,SAAS,gBAAgB;AAClE,cAAM,IAAI;AAAA,UACR,qCAAqC,KAAK,UAAU;AAAA,UACpD,8EAA8E,KAAK,QAAQ;AAAA,QAC7F;AAAA,MACF;AACA,YAAM,IAAI;AAAA,QACR,0CAA0C,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK,CAAC;AAAA,QAChG,uCAAuC,KAAK,QAAQ;AAAA,MACtD;AAAA,IACF;AAEA,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,YAAY,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACtD,YAAM,IAAI;AAAA,QACR,0BAA0B,SAAS,MAAM,KAAK,SAAS;AAAA,QACvD,SAAS,WAAW,MAChB,yDACA,yCAAyC,GAAG;AAAA,MAClD;AAAA,IACF;AAEA,WAAQ,MAAM,SAAS,KAAK;AAAA,EAC9B;AAAA,EAEQ,mBAAmB,QAAwB;AACjD,QAAI,KAAK,gBAAgB,QAAW;AAClC,WAAK,cAAc,OAAO;AAAA,IAC5B;AAAA,EACF;AACF;AAGA,SAAS,iBAAiB,GAAa,GAAqB;AAC1D,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC;AACd,UAAM,KAAK,EAAE,CAAC;AACd,WAAO,KAAK;AACZ,aAAS,KAAK;AACd,aAAS,KAAK;AAAA,EAChB;AACA,QAAM,QAAQ,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,KAAK;AAChD,SAAO,UAAU,IAAI,IAAI,MAAM;AACjC;","names":[]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@utaba/deep-memory-embeddings-openai",
-  "version": "0.8.1",
+  "version": "0.9.0",
   "description": "OpenAI-compatible embeddings provider for @utaba/deep-memory",
   "type": "module",
   "main": "./dist/index.cjs",
@@ -52,7 +52,7 @@
     "agents"
   ],
   "dependencies": {
-    "@utaba/deep-memory": "0.8.1"
+    "@utaba/deep-memory": "0.9.0"
   },
   "devDependencies": {
     "@types/node": "^22.0.0",
@@ -61,7 +61,7 @@
     "vitest": "^4.1.2"
   },
   "peerDependencies": {
-    "@utaba/deep-memory": "^0.8.1"
+    "@utaba/deep-memory": "^0.9.0"
   },
   "scripts": {
     "build": "tsup",