@superlinked/sie-langchain 0.1.9 → 0.1.10

package/dist/index.cjs

@@ -21,11 +21,192 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 var index_exports = {};
 __export(index_exports, {
   SIEEmbeddings: () => SIEEmbeddings,
+  SIEExtractor: () => SIEExtractor,
+  SIEReranker: () => SIEReranker,
   SIESparseEncoder: () => SIESparseEncoder
 });
 module.exports = __toCommonJS(index_exports);
 var import_embeddings = require("@langchain/core/embeddings");
+var import_sie_sdk3 = require("@superlinked/sie-sdk");
+
+// src/rerankers.ts
+var import_document_compressors = require("@langchain/core/retrievers/document_compressors");
 var import_sie_sdk = require("@superlinked/sie-sdk");
+var SIEReranker = class extends import_document_compressors.BaseDocumentCompressor {
+  model;
+  topK;
+  _client;
+  _ownsClient;
+  baseUrl;
+  clientOptions;
+  constructor(params = {}) {
+    super();
+    const {
+      baseUrl = "http://localhost:8080",
+      model = "jinaai/jina-reranker-v2-base-multilingual",
+      client,
+      topK,
+      gpu,
+      timeout = 18e4
+    } = params;
+    this.baseUrl = baseUrl;
+    this.model = model;
+    this.topK = topK;
+    this._client = client;
+    this._ownsClient = !client;
+    this.clientOptions = {
+      timeout,
+      gpu
+    };
+  }
+  /**
+   * Get or create the SIEClient.
+   */
+  get client() {
+    if (!this._client) {
+      this._client = new import_sie_sdk.SIEClient(this.baseUrl, this.clientOptions);
+    }
+    return this._client;
+  }
+  /**
+   * Rerank documents by relevance to query.
+   *
+   * @param documents - Documents to rerank.
+   * @param query - Query to rank documents against.
+   * @returns Reranked documents with relevance_score in metadata, sorted by score descending.
+   */
+  async compressDocuments(documents, query) {
+    if (documents.length === 0) {
+      return [];
+    }
+    const queryItem = { text: query };
+    const docItems = documents.map((doc) => ({ text: doc.pageContent }));
+    const result = await this.client.score(this.model, queryItem, docItems);
+    const reranked = [];
+    for (const entry of result.scores) {
+      const idx = Number.parseInt(entry.itemId, 10);
+      const doc = documents[idx];
+      if (doc) {
+        reranked.push({
+          pageContent: doc.pageContent,
+          metadata: { ...doc.metadata, relevance_score: entry.score },
+          id: doc.id
+        });
+      }
+    }
+    if (this.topK !== void 0) {
+      return reranked.slice(0, this.topK);
+    }
+    return reranked;
+  }
+  /**
+   * Close the underlying client connection.
+   */
+  async close() {
+    if (this._client && this._ownsClient) {
+      await this._client.close();
+    }
+  }
+};
+
+// src/extractors.ts
+var import_tools = require("@langchain/core/tools");
+var import_sie_sdk2 = require("@superlinked/sie-sdk");
+var SIEExtractor = class extends import_tools.Tool {
+  name;
+  description;
+  model;
+  labels;
+  threshold;
+  _client;
+  _ownsClient;
+  baseUrl;
+  clientOptions;
+  constructor(params = {}) {
+    const toolName = params.name ?? "sie_extract";
+    const toolDescription = params.description ?? "Extract structured information from text. Input should be text to analyze. Returns JSON with entities, relations, classifications, and detected objects.";
+    super({});
+    this.name = toolName;
+    this.description = toolDescription;
+    const {
+      baseUrl = "http://localhost:8080",
+      model = "urchade/gliner_multi-v2.1",
+      client,
+      labels = ["person", "organization", "location"],
+      threshold,
+      gpu,
+      timeout = 18e4
+    } = params;
+    this.baseUrl = baseUrl;
+    this.model = model;
+    this.labels = labels;
+    this.threshold = threshold;
+    this._client = client;
+    this._ownsClient = !client;
+    this.clientOptions = {
+      timeout,
+      gpu
+    };
+  }
+  /**
+   * Get or create the SIEClient.
+   */
+  get client() {
+    if (!this._client) {
+      this._client = new import_sie_sdk2.SIEClient(this.baseUrl, this.clientOptions);
+    }
+    return this._client;
+  }
+  /**
+   * Extract structured information from text.
+   *
+   * @param text - Text to extract from.
+   * @returns JSON string with entities, relations, classifications, and objects.
+   */
+  async _call(text) {
+    const extractOptions = {
+      labels: this.labels
+    };
+    if (this.threshold !== void 0) {
+      extractOptions.threshold = this.threshold;
+    }
+    const result = await this.client.extract(this.model, { text }, extractOptions);
+    return JSON.stringify({
+      entities: result.entities.map((e) => ({
+        text: e.text,
+        label: e.label,
+        score: e.score,
+        ...e.start !== void 0 && { start: e.start },
+        ...e.end !== void 0 && { end: e.end }
+      })),
+      relations: result.relations.map((r) => ({
+        head: r.head,
+        tail: r.tail,
+        relation: r.relation,
+        score: r.score
+      })),
+      classifications: result.classifications.map((c) => ({
+        label: c.label,
+        score: c.score
+      })),
+      objects: result.objects.map((o) => ({
+        label: o.label,
+        score: o.score,
+        bbox: o.bbox
+      }))
+    });
+  }
+  /**
+   * Close the underlying client connection.
+   */
+  async close() {
+    if (this._client && this._ownsClient) {
+      await this._client.close();
+    }
+  }
+};
+
+// src/index.ts
 var SIEEmbeddings = class extends import_embeddings.Embeddings {
   model;
   instruction;
@@ -63,7 +244,7 @@ var SIEEmbeddings = class extends import_embeddings.Embeddings {
   get client() {
     if (!this._client) {
       const baseUrl = this.baseUrl ?? "http://localhost:8080";
-      this._client = new import_sie_sdk.SIEClient(baseUrl, this.clientOptions);
+      this._client = new import_sie_sdk3.SIEClient(baseUrl, this.clientOptions);
     }
     return this._client;
   }
@@ -85,7 +266,7 @@ var SIEEmbeddings = class extends import_embeddings.Embeddings {
       isQuery: false
     };
     const results = await this.client.encode(this.model, items, options);
-    return results.map((result) => this.extractDense(result));
+    return results.map((result) => (0, import_sie_sdk3.denseEmbedding)(result));
   }
   /**
    * Embed a single query text.
@@ -103,17 +284,7 @@ var SIEEmbeddings = class extends import_embeddings.Embeddings {
       isQuery: true
     };
     const result = await this.client.encode(this.model, { text }, options);
-    return this.extractDense(result);
-  }
-  /**
-   * Extract dense embedding from encode result.
-   */
-  extractDense(result) {
-    const dense = result.dense;
-    if (!dense) {
-      throw new Error("Encode result missing dense embedding");
-    }
-    return (0, import_sie_sdk.toNumberArray)(dense);
+    return (0, import_sie_sdk3.denseEmbedding)(result);
   }
   /**
    * Close the underlying client connection.
@@ -148,7 +319,7 @@ var SIESparseEncoder = class {
    */
   get client() {
     if (!this._client) {
-      this._client = new import_sie_sdk.SIEClient(this.baseUrl, this.clientOptions);
+      this._client = new import_sie_sdk3.SIEClient(this.baseUrl, this.clientOptions);
     }
     return this._client;
   }
@@ -168,7 +339,7 @@ var SIESparseEncoder = class {
       isQuery: true
     };
     const results = await this.client.encode(this.model, items, options);
-    return results.map((result) => this.extractSparse(result));
+    return results.map((result) => (0, import_sie_sdk3.sparseEmbedding)(result));
   }
   /**
    * Encode document texts to sparse vectors.
@@ -186,20 +357,7 @@ var SIESparseEncoder = class {
       isQuery: false
     };
     const results = await this.client.encode(this.model, items, options);
-    return results.map((result) => this.extractSparse(result));
-  }
-  /**
-   * Extract sparse embedding from encode result.
-   */
-  extractSparse(result) {
-    const sparse = result.sparse;
-    if (!sparse) {
-      return { indices: [], values: [] };
-    }
-    return {
-      indices: (0, import_sie_sdk.toNumberArray)(sparse.indices),
-      values: (0, import_sie_sdk.toNumberArray)(sparse.values)
-    };
+    return results.map((result) => (0, import_sie_sdk3.sparseEmbedding)(result));
   }
   /**
    * Close the underlying client connection.
@@ -213,6 +371,8 @@ var SIESparseEncoder = class {
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   SIEEmbeddings,
+  SIEExtractor,
+  SIEReranker,
   SIESparseEncoder
 });
 //# sourceMappingURL=index.cjs.map

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts"],"sourcesContent":["/**\n * SIE embeddings integration for LangChain.js\n *\n * Provides drop-in replacement for OpenAI embeddings using SIE's inference server:\n * - SIEEmbeddings: Dense embeddings for vector stores\n * - SIESparseEncoder: Sparse encoder for hybrid search\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * const vectors = await embeddings.embedDocuments([\"Hello world\"]);\n * const queryVector = await embeddings.embedQuery(\"What is hello?\");\n * ```\n */\n\nimport { Embeddings, type EmbeddingsParams } from \"@langchain/core/embeddings\";\nimport {\n  type DType,\n  type EncodeOptions,\n  type EncodeResult,\n  SIEClient,\n  type SIEClientOptions,\n  toNumberArray,\n} from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEEmbeddings.\n */\nexport interface SIEEmbeddingsParams extends EmbeddingsParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Optional instruction prefix for embedding (model-dependent).\n   */\n  instruction?: string;\n\n  /**\n   * Output dtype: \"float32\" (default), \"float16\", \"int8\", \"binary\".\n   */\n  outputDtype?: DType;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * LangChain Embeddings implementation using SIE.\n *\n * Wraps SIEClient.encode() to implement the LangChain Embeddings interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * // Basic usage\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * // Embed documents\n * const docVectors = await embeddings.embedDocuments([\n *   \"First document\",\n *   \"Second document\",\n * ]);\n *\n * // Embed a query (may use different encoding for asymmetric models)\n * const queryVector = await embeddings.embedQuery(\"What is the topic?\");\n *\n * // With GPU routing\n * const gpuEmbeddings = new SIEEmbeddings({\n *   baseUrl: \"https://cluster.example.com\",\n *   model: \"BAAI/bge-m3\",\n *   gpu: \"a100-80gb\",\n * });\n * ```\n */\nexport class SIEEmbeddings extends Embeddings {\n  private readonly model: string;\n  private readonly instruction?: string;\n  private readonly outputDtype?: DType;\n  private _client: SIEClient | undefined;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIEEmbeddingsParams = {}) {\n    super(params);\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      client,\n      instruction,\n      outputDtype,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.model = model;\n    this.instruction = instruction;\n    this.outputDtype = outputDtype;\n    this._client = client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n\n    // If no client provided, we'll create one lazily using baseUrl\n    if (!client) {\n      this.clientOptions.timeout = timeout;\n      this.clientOptions.gpu = gpu;\n      // Store baseUrl for lazy client creation\n      (this as { baseUrl?: string }).baseUrl = baseUrl;\n    }\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      const baseUrl = (this as { baseUrl?: string }).baseUrl ?? \"http://localhost:8080\";\n      this._client = new SIEClient(baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Embed a list of documents.\n   *\n   * @param texts - List of document texts to embed.\n   * @returns List of embedding vectors (as arrays of numbers).\n   */\n  async embedDocuments(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => this.extractDense(result));\n  }\n\n  /**\n   * Embed a single query text.\n   *\n   * For asymmetric models (like BGE-M3), this uses query-specific encoding.\n   *\n   * @param text - Query text to embed.\n   * @returns Embedding vector as array of numbers.\n   */\n  async embedQuery(text: string): Promise<number[]> {\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: true,\n    };\n\n    const result = await this.client.encode(this.model, { text }, options);\n    return this.extractDense(result as EncodeResult);\n  }\n\n  /**\n   * Extract dense embedding from encode result.\n   */\n  private extractDense(result: EncodeResult): number[] {\n    const dense = result.dense;\n    if (!dense) {\n      throw new Error(\"Encode result missing dense embedding\");\n    }\n    return toNumberArray(dense);\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n\n/**\n * Configuration options for SIESparseEncoder.\n */\nexport interface SIESparseEncoderOptions {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding. Must support sparse output.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * Sparse encoder for LangChain hybrid search.\n *\n * Compatible with PineconeHybridSearchRetriever's sparse_encoder interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings, SIESparseEncoder } from \"@superlinked/sie-langchain\";\n * import { PineconeHybridSearchRetriever } from \"@langchain/pinecone\";\n *\n * const retriever = new PineconeHybridSearchRetriever({\n *   embeddings: new SIEEmbeddings({ model: \"BAAI/bge-m3\" }),\n *   sparseEncoder: new SIESparseEncoder({ model: \"BAAI/bge-m3\" }),\n *   index: pineconeIndex,\n * });\n * ```\n */\nexport class SIESparseEncoder {\n  private readonly model: string;\n  private _client: SIEClient | undefined;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(options: SIESparseEncoderOptions = {}) {\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      gpu,\n      timeout = 180_000,\n    } = options;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Encode query texts to sparse vectors.\n   *\n   * @param texts - List of query texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeQueries(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: true,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => this.extractSparse(result));\n  }\n\n  /**\n   * Encode document texts to sparse vectors.\n   *\n   * @param texts - List of document texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeDocuments(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => this.extractSparse(result));\n  }\n\n  /**\n   * Extract sparse embedding from encode result.\n   */\n  private extractSparse(result: EncodeResult): { indices: number[]; values: number[] } {\n    const sparse = result.sparse;\n    if (!sparse) {\n      return { indices: [], values: [] };\n    }\n\n    return {\n      indices: toNumberArray(sparse.indices),\n      values: toNumberArray(sparse.values),\n    };\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAqBA,wBAAkD;AAClD,qBAOO;AA8EA,IAAM,gBAAN,cAA4B,6BAAW;AAAA,EAC3B;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EAEjB,YAAY,SAA8B,CAAC,GAAG;AAC5C,UAAM,MAAM;AAEZ,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,QAAQ;AACb,SAAK,cAAc;AACnB,SAAK,cAAc;AACnB,SAAK,UAAU;AAEf,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAGA,QAAI,CAAC,QAAQ;AACX,WAAK,cAAc,UAAU;AAC7B,WAAK,cAAc,MAAM;AAEzB,MAAC,KAA8B,UAAU;AAAA,IAC3C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,YAAM,UAAW,KAA8B,WAAW;AAC1D,WAAK,UAAU,IAAI,yBAAU,SAAS,KAAK,aAAa;AAAA,IAC1D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eAAe,OAAsC;AACzD,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,KAAK,aAAa,MAAM,CAAC;AAAA,EAC9E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,MAAM,WAAW,MAAiC;AAChD,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,SAAS,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,EAAE,KAAK,GAAG,OAAO;AACrE,WAAO,KAAK,aAAa,MAAsB;AAAA,EACjD;AAAA;AAAA;AAAA;AAAA,EAKQ,aAAa,QAAgC;AACnD,UAAM,QAAQ,OAAO;AACrB,QAAI,CAAC,OAAO;AACV,YAAM,IAAI,MAAM,uCAAuC;AAAA,IACzD;AACA,eAAO,8BAAc,KAAK;AAAA,EAC5B;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;AA+CO,IAAM,mBAAN,MAAuB;AAAA,EACX;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EAEjB,YAAY,UAAmC,CAAC,GAAG;AACjD,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAI,yBAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,cAAc,OAA0E;AAC5F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,KAAK,cAAc,MAAM,CAAC;AAAA,EAC/E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,gBAAgB,OAA0E;AAC9F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,KAAK,cAAc,MAAM,CAAC;AAAA,EAC/E;AAAA;AAAA;AAAA;AAAA,EAKQ,cAAc,QAA+D;AACnF,UAAM,SAAS,OAAO;AACtB,QAAI,CAAC,QAAQ;AACX,aAAO,EAAE,SAAS,CAAC,GAAG,QAAQ,CAAC,EAAE;AAAA,IACnC;AAEA,WAAO;AAAA,MACL,aAAS,8BAAc,OAAO,OAAO;AAAA,MACrC,YAAQ,8BAAc,OAAO,MAAM;AAAA,IACrC;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;","names":[]}
+{"version":3,"sources":["../src/index.ts","../src/rerankers.ts","../src/extractors.ts"],"sourcesContent":["/**\n * SIE embeddings integration for LangChain.js\n *\n * Provides drop-in replacement for OpenAI embeddings using SIE's inference server:\n * - SIEEmbeddings: Dense embeddings for vector stores\n * - SIESparseEncoder: Sparse encoder for hybrid search\n * - SIEReranker: Cross-encoder reranking for retrieval pipelines\n * - SIEExtractor: Entity extraction tool for agents\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * const vectors = await embeddings.embedDocuments([\"Hello world\"]);\n * const queryVector = await embeddings.embedQuery(\"What is hello?\");\n * ```\n */\n\nimport { Embeddings, type EmbeddingsParams } from \"@langchain/core/embeddings\";\nimport {\n  type DType,\n  type EncodeOptions,\n  type EncodeResult,\n  SIEClient,\n  type SIEClientOptions,\n  denseEmbedding,\n  sparseEmbedding,\n} from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEEmbeddings.\n */\nexport interface SIEEmbeddingsParams extends EmbeddingsParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Optional instruction prefix for embedding (model-dependent).\n   */\n  instruction?: string;\n\n  /**\n   * Output dtype: \"float32\" (default), \"float16\", \"int8\", \"binary\".\n   */\n  outputDtype?: DType;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * LangChain Embeddings implementation using SIE.\n *\n * Wraps SIEClient.encode() to implement the LangChain Embeddings interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * // Basic usage\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * // Embed documents\n * const docVectors = await embeddings.embedDocuments([\n *   \"First document\",\n *   \"Second document\",\n * ]);\n *\n * // Embed a query (may use different encoding for asymmetric models)\n * const queryVector = await embeddings.embedQuery(\"What is the topic?\");\n *\n * // With GPU routing\n * const gpuEmbeddings = new SIEEmbeddings({\n *   baseUrl: \"https://cluster.example.com\",\n *   model: \"BAAI/bge-m3\",\n *   gpu: \"a100-80gb\",\n * });\n * ```\n */\nexport class SIEEmbeddings extends Embeddings {\n  private readonly model: string;\n  private readonly instruction?: string;\n  private readonly outputDtype?: DType;\n  private _client: SIEClient | undefined;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIEEmbeddingsParams = {}) {\n    super(params);\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      client,\n      instruction,\n      outputDtype,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.model = model;\n    this.instruction = instruction;\n    this.outputDtype = outputDtype;\n    this._client = client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n\n    // If no client provided, we'll create one lazily using baseUrl\n    if (!client) {\n      this.clientOptions.timeout = timeout;\n      this.clientOptions.gpu = gpu;\n      // Store baseUrl for lazy client creation\n      (this as { baseUrl?: string }).baseUrl = baseUrl;\n    }\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      const baseUrl = (this as { baseUrl?: string }).baseUrl ?? \"http://localhost:8080\";\n      this._client = new SIEClient(baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Embed a list of documents.\n   *\n   * @param texts - List of document texts to embed.\n   * @returns List of embedding vectors (as arrays of numbers).\n   */\n  async embedDocuments(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => denseEmbedding(result));\n  }\n\n  /**\n   * Embed a single query text.\n   *\n   * For asymmetric models (like BGE-M3), this uses query-specific encoding.\n   *\n   * @param text - Query text to embed.\n   * @returns Embedding vector as array of numbers.\n   */\n  async embedQuery(text: string): Promise<number[]> {\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: true,\n    };\n\n    const result = await this.client.encode(this.model, { text }, options);\n    return denseEmbedding(result as EncodeResult);\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n\n/**\n * Configuration options for SIESparseEncoder.\n */\nexport interface SIESparseEncoderOptions {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding. Must support sparse output.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * Sparse encoder for LangChain hybrid search.\n *\n * Compatible with PineconeHybridSearchRetriever's sparse_encoder interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings, SIESparseEncoder } from \"@superlinked/sie-langchain\";\n * import { PineconeHybridSearchRetriever } from \"@langchain/pinecone\";\n *\n * const retriever = new PineconeHybridSearchRetriever({\n *   embeddings: new SIEEmbeddings({ model: \"BAAI/bge-m3\" }),\n *   sparseEncoder: new SIESparseEncoder({ model: \"BAAI/bge-m3\" }),\n *   index: pineconeIndex,\n * });\n * ```\n */\nexport class SIESparseEncoder {\n  private readonly model: string;\n  private _client: SIEClient | undefined;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(options: SIESparseEncoderOptions = {}) {\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      gpu,\n      timeout = 180_000,\n    } = options;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Encode query texts to sparse vectors.\n   *\n   * @param texts - List of query texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeQueries(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: true,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => sparseEmbedding(result));\n  }\n\n  /**\n   * Encode document texts to sparse vectors.\n   *\n   * @param texts - List of document texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeDocuments(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => sparseEmbedding(result));\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n\nexport { SIEReranker, type SIERerankerParams } from \"./rerankers.js\";\nexport { SIEExtractor, type SIEExtractorParams } from \"./extractors.js\";\n","/**\n * SIE reranker integration for LangChain.js\n *\n * Provides document reranking using SIE's score endpoint:\n * - SIEReranker: Cross-encoder reranking implementing BaseDocumentCompressor\n *\n * @example\n * ```typescript\n * import { SIEReranker } from \"@superlinked/sie-langchain\";\n *\n * const reranker = new SIEReranker({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"jinaai/jina-reranker-v2-base-multilingual\",\n *   topK: 3,\n * });\n *\n * const reranked = await reranker.compressDocuments(documents, \"search query\");\n * ```\n */\n\nimport type { DocumentInterface } from \"@langchain/core/documents\";\nimport { BaseDocumentCompressor } from \"@langchain/core/retrievers/document_compressors\";\nimport { SIEClient, type SIEClientOptions } from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEReranker.\n */\nexport interface SIERerankerParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Reranker model name/ID.\n   * @default \"jinaai/jina-reranker-v2-base-multilingual\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Number of top documents to return. If undefined, returns all documents.\n   */\n  topK?: number;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * LangChain document compressor using SIE's reranking.\n *\n * Wraps SIEClient.score() to implement BaseDocumentCompressor.\n *\n * @example\n * ```typescript\n * import { SIEReranker } from \"@superlinked/sie-langchain\";\n *\n * const reranker = new SIEReranker({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"jinaai/jina-reranker-v2-base-multilingual\",\n *   topK: 3,\n * });\n *\n * // Rerank retrieved documents\n * const reranked = await reranker.compressDocuments(documents, \"search query\");\n *\n * // Use in a retrieval pipeline\n * import { ContextualCompressionRetriever } from \"langchain/retrievers/contextual_compression\";\n *\n * const compressionRetriever = new ContextualCompressionRetriever({\n *   baseCompressor: reranker,\n *   baseRetriever: vectorStoreRetriever,\n * });\n * ```\n */\nexport class SIEReranker extends BaseDocumentCompressor {\n  private readonly model: string;\n  private readonly topK?: number;\n  private _client: SIEClient | undefined;\n  private readonly _ownsClient: boolean;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIERerankerParams = {}) {\n    super();\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"jinaai/jina-reranker-v2-base-multilingual\",\n      client,\n      topK,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.topK = topK;\n    this._client = client;\n    this._ownsClient = !client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Rerank documents by relevance to query.\n   *\n   * @param documents - Documents to rerank.\n   * @param query - Query to rank documents against.\n   * @returns Reranked documents with relevance_score in metadata, sorted by score descending.\n   */\n  async compressDocuments(\n    documents: DocumentInterface[],\n    query: string,\n  ): Promise<DocumentInterface[]> {\n    if (documents.length === 0) {\n      return [];\n    }\n\n    const queryItem = { text: query };\n    const docItems = documents.map((doc) => ({ text: doc.pageContent }));\n\n    const result = await this.client.score(this.model, queryItem, docItems);\n\n    // Map score entries back to documents with relevance_score in metadata.\n    // ScoreResult.scores are already sorted by score descending.\n    const reranked: DocumentInterface[] = [];\n    for (const entry of result.scores) {\n      const idx = Number.parseInt(entry.itemId, 10);\n      const doc = documents[idx];\n      if (doc) {\n        reranked.push({\n          pageContent: doc.pageContent,\n          metadata: { ...doc.metadata, relevance_score: entry.score },\n          id: doc.id,\n        });\n      }\n    }\n\n    if (this.topK !== undefined) {\n      return reranked.slice(0, this.topK);\n    }\n    return reranked;\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client && this._ownsClient) {\n      await this._client.close();\n    }\n  }\n}\n","/**\n * SIE extraction tool for LangChain.js\n *\n * Provides extraction using SIE's extract endpoint:\n * - SIEExtractor: Extraction tool implementing LangChain Tool\n *\n * Returns entities, relations, classifications, and detected objects.\n *\n * @example\n * ```typescript\n * import { SIEExtractor } from \"@superlinked/sie-langchain\";\n *\n * const extractor = new SIEExtractor({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"urchade/gliner_multi-v2.1\",\n *   labels: [\"person\", \"organization\", \"location\"],\n * });\n *\n * const result = await extractor.invoke(\"John Smith works at Acme Corp in NYC\");\n * const parsed = JSON.parse(result);\n * console.log(parsed.entities);\n * console.log(parsed.relations);\n * ```\n */\n\nimport { Tool } from \"@langchain/core/tools\";\nimport {\n  type ExtractOptions,\n  type ExtractResult,\n  SIEClient,\n  type SIEClientOptions,\n} from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEExtractor.\n */\nexport interface SIEExtractorParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Extraction model name/ID.\n   * @default \"urchade/gliner_multi-v2.1\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Labels to extract (entity types, relation types, or classification labels).\n   * @default [\"person\", \"organization\", \"location\"]\n   */\n  labels?: string[];\n\n  /**\n   * Minimum confidence threshold (0-1).\n   */\n  threshold?: number;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n\n  /**\n   * Tool name for use in agents.\n   * @default \"sie_extract\"\n   */\n  name?: string;\n\n  /**\n   * Tool description for use in agents.\n   */\n  description?: string;\n}\n\n/**\n * LangChain tool for extraction using SIE.\n *\n * Wraps SIEClient.extract() to implement the LangChain Tool interface\n * for use in agents and chains. Returns JSON with entities, relations,\n * classifications, and detected objects.\n *\n * @example\n * ```typescript\n * import { SIEExtractor } from \"@superlinked/sie-langchain\";\n *\n * // Direct usage\n * const extractor = new SIEExtractor({\n *   model: \"urchade/gliner_multi-v2.1\",\n *   labels: [\"person\", \"organization\", \"location\"],\n * });\n * const result = await extractor.invoke(\"John Smith works at Acme Corp\");\n * const parsed = JSON.parse(result);\n *\n * // Use in an agent\n * import { ChatOpenAI } from \"@langchain/openai\";\n * import { createReactAgent } from \"@langchain/langgraph/prebuilt\";\n *\n * const agent = createReactAgent({\n *   llm: new ChatOpenAI(),\n *   tools: [extractor],\n * });\n * ```\n */\nexport class SIEExtractor extends Tool {\n  name: string;\n  description: string;\n\n  private readonly model: string;\n  private readonly labels: string[];\n  private readonly threshold?: number;\n  private _client: SIEClient | undefined;\n  private readonly _ownsClient: boolean;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIEExtractorParams = {}) {\n    const toolName = params.name ?? \"sie_extract\";\n    const toolDescription =\n      params.description ??\n      \"Extract structured information from text. \" +\n        \"Input should be text to analyze. \" +\n        \"Returns JSON with entities, relations, classifications, and detected objects.\";\n\n    super({});\n\n    this.name = toolName;\n    this.description = toolDescription;\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"urchade/gliner_multi-v2.1\",\n      client,\n      labels = [\"person\", \"organization\", \"location\"],\n      threshold,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.labels = labels;\n    this.threshold = threshold;\n    this._client = client;\n    this._ownsClient = !client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Extract structured information from text.\n   *\n   * @param text - Text to extract from.\n   * @returns JSON string with entities, relations, classifications, and objects.\n   */\n  async _call(text: string): Promise<string> {\n    const extractOptions: ExtractOptions = {\n      labels: this.labels,\n    };\n    if (this.threshold !== undefined) {\n      extractOptions.threshold = this.threshold;\n    }\n\n    const result: ExtractResult = await this.client.extract(this.model, { text }, extractOptions);\n\n    return JSON.stringify({\n      entities: result.entities.map((e) => ({\n        text: e.text,\n        label: e.label,\n        score: e.score,\n        ...(e.start !== undefined && { start: e.start }),\n        ...(e.end !== undefined && { end: e.end }),\n      })),\n      relations: result.relations.map((r) => ({\n        head: r.head,\n        tail: r.tail,\n        relation: r.relation,\n        score: r.score,\n      })),\n      classifications: result.classifications.map((c) => ({\n        label: c.label,\n        score: c.score,\n      })),\n      objects: result.objects.map((o) => ({\n        label: o.label,\n        score: o.score,\n        bbox: o.bbox,\n      })),\n    });\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client && this._ownsClient) {\n      await this._client.close();\n    }\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAuBA,wBAAkD;AAClD,IAAAA,kBAQO;;;ACXP,kCAAuC;AACvC,qBAAiD;AAoE1C,IAAM,cAAN,cAA0B,mDAAuB;AAAA,EACrC;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EACA;AAAA,EAEjB,YAAY,SAA4B,CAAC,GAAG;AAC1C,UAAM;AAEN,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,OAAO;AACZ,SAAK,UAAU;AACf,SAAK,cAAc,CAAC;AAEpB,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAI,yBAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,kBACJ,WACA,OAC8B;AAC9B,QAAI,UAAU,WAAW,GAAG;AAC1B,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,YAAY,EAAE,MAAM,MAAM;AAChC,UAAM,WAAW,UAAU,IAAI,CAAC,SAAS,EAAE,MAAM,IAAI,YAAY,EAAE;AAEnE,UAAM,SAAS,MAAM,KAAK,OAAO,MAAM,KAAK,OAAO,WAAW,QAAQ;AAItE,UAAM,WAAgC,CAAC;AACvC,eAAW,SAAS,OAAO,QAAQ;AACjC,YAAM,MAAM,OAAO,SAAS,MAAM,QAAQ,EAAE;AAC5C,YAAM,MAAM,UAAU,GAAG;AACzB,UAAI,KAAK;AACP,iBAAS,KAAK;AAAA,UACZ,aAAa,IAAI;AAAA,UACjB,UAAU,EAAE,GAAG,IAAI,UAAU,iBAAiB,MAAM,MAAM;AAAA,UAC1D,IAAI,IAAI;AAAA,QACV,CAAC;AAAA,MACH;AAAA,IACF;AAEA,QAAI,KAAK,SAAS,QAAW;AAC3B,aAAO,SAAS,MAAM,GAAG,KAAK,IAAI;AAAA,IACpC;AACA,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,WAAW,KAAK,aAAa;AACpC,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;;;AC5JA,mBAAqB;AACrB,IAAAC,kBAKO;AAuFA,IAAM,eAAN,cAA2B,kBAAK;AAAA,EACrC;AAAA,EACA;AAAA,EAEiB;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EACA;AAAA,EAEjB,YAAY,SAA6B,CAAC,GAAG;AAC3C,UAAM,WAAW,OAAO,QAAQ;AAChC,UAAM,kBACJ,OAAO,eACP;AAIF,UAAM,CAAC,CAAC;AAER,SAAK,OAAO;AACZ,SAAK,cAAc;AAEnB,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA,SAAS,CAAC,UAAU,gBAAgB,UAAU;AAAA,MAC9C;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,YAAY;AACjB,SAAK,UAAU;AACf,SAAK,cAAc,CAAC;AAEpB,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAI,0BAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,MAAM,MAA+B;AACzC,UAAM,iBAAiC;AAAA,MACrC,QAAQ,KAAK;AAAA,IACf;AACA,QAAI,KAAK,cAAc,QAAW;AAChC,qBAAe,YAAY,KAAK;AAAA,IAClC;AAEA,UAAM,SAAwB,MAAM,KAAK,OAAO,QAAQ,KAAK,OAAO,EAAE,KAAK,GAAG,cAAc;AAE5F,WAAO,KAAK,UAAU;AAAA,MACpB,UAAU,OAAO,SAAS,IAAI,CAAC,OAAO;AAAA,QACpC,MAAM,EAAE;AAAA,QACR,OAAO,EAAE;AAAA,QACT,OAAO,EAAE;AAAA,QACT,GAAI,EAAE,UAAU,UAAa,EAAE,OAAO,EAAE,MAAM;AAAA,QAC9C,GAAI,EAAE,QAAQ,UAAa,EAAE,KAAK,EAAE,IAAI;AAAA,MAC1C,EAAE;AAAA,MACF,WAAW,OAAO,UAAU,IAAI,CAAC,OAAO;AAAA,QACtC,MAAM,EAAE;AAAA,QACR,MAAM,EAAE;AAAA,QACR,UAAU,EAAE;AAAA,QACZ,OAAO,EAAE;AAAA,MACX,EAAE;AAAA,MACF,iBAAiB,OAAO,gBAAgB,IAAI,CAAC,OAAO;AAAA,QAClD,OAAO,EAAE;AAAA,QACT,OAAO,EAAE;AAAA,MACX,EAAE;AAAA,MACF,SAAS,OAAO,QAAQ,IAAI,CAAC,OAAO;AAAA,QAClC,OAAO,EAAE;AAAA,QACT,OAAO,EAAE;AAAA,QACT,MAAM,EAAE;AAAA,MACV,EAAE;AAAA,IACJ,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,WAAW,KAAK,aAAa;AACpC,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;;;AFpHO,IAAM,gBAAN,cAA4B,6BAAW;AAAA,EAC3B;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EAEjB,YAAY,SAA8B,CAAC,GAAG;AAC5C,UAAM,MAAM;AAEZ,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,QAAQ;AACb,SAAK,cAAc;AACnB,SAAK,cAAc;AACnB,SAAK,UAAU;AAEf,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAGA,QAAI,CAAC,QAAQ;AACX,WAAK,cAAc,UAAU;AAC7B,WAAK,cAAc,MAAM;AAEzB,MAAC,KAA8B,UAAU;AAAA,IAC3C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,YAAM,UAAW,KAA8B,WAAW;AAC1D,WAAK,UAAU,IAAI,0BAAU,SAAS,KAAK,aAAa;AAAA,IAC1D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eAAe,OAAsC;AACzD,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,eAAW,gCAAe,MAAM,CAAC;AAAA,EAC3E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,MAAM,WAAW,MAAiC;AAChD,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,SAAS,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,EAAE,KAAK,GAAG,OAAO;AACrE,eAAO,gCAAe,MAAsB;AAAA,EAC9C;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;AA+CO,IAAM,mBAAN,MAAuB;AAAA,EACX;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EAEjB,YAAY,UAAmC,CAAC,GAAG;AACjD,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAI,0BAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,cAAc,OAA0E;AAC5F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,eAAW,iCAAgB,MAAM,CAAC;AAAA,EAC5E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,gBAAgB,OAA0E;AAC9F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,eAAW,iCAAgB,MAAM,CAAC;AAAA,EAC5E;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;","names":["import_sie_sdk","import_sie_sdk"]}

package/dist/index.d.cts

@@ -1,5 +1,243 @@
 import { EmbeddingsParams, Embeddings } from '@langchain/core/embeddings';
 import { SIEClient, DType } from '@superlinked/sie-sdk';
+import { DocumentInterface } from '@langchain/core/documents';
+import { BaseDocumentCompressor } from '@langchain/core/retrievers/document_compressors';
+import { Tool } from '@langchain/core/tools';
+
+/**
+ * SIE reranker integration for LangChain.js
+ *
+ * Provides document reranking using SIE's score endpoint:
+ * - SIEReranker: Cross-encoder reranking implementing BaseDocumentCompressor
+ *
+ * @example
+ * ```typescript
+ * import { SIEReranker } from "@superlinked/sie-langchain";
+ *
+ * const reranker = new SIEReranker({
+ *   baseUrl: "http://localhost:8080",
+ *   model: "jinaai/jina-reranker-v2-base-multilingual",
+ *   topK: 3,
+ * });
+ *
+ * const reranked = await reranker.compressDocuments(documents, "search query");
+ * ```
+ */
+
+/**
+ * Configuration options for SIEReranker.
+ */
+interface SIERerankerParams {
+    /**
+     * URL of the SIE server.
+     * @default "http://localhost:8080"
+     */
+    baseUrl?: string;
+    /**
+     * Reranker model name/ID.
+     * @default "jinaai/jina-reranker-v2-base-multilingual"
+     */
+    model?: string;
+    /**
+     * Optional pre-configured SIEClient instance.
+     * If provided, baseUrl and other connection options are ignored.
+     */
+    client?: SIEClient;
+    /**
+     * Number of top documents to return. If undefined, returns all documents.
+     */
+    topK?: number;
+    /**
+     * Target GPU type for routing (e.g., "l4", "a100-80gb").
+     */
+    gpu?: string;
+    /**
+     * Request timeout in milliseconds.
+     * @default 180000 (3 minutes)
+     */
+    timeout?: number;
+}
+/**
+ * LangChain document compressor using SIE's reranking.
+ *
+ * Wraps SIEClient.score() to implement BaseDocumentCompressor.
+ *
+ * @example
+ * ```typescript
+ * import { SIEReranker } from "@superlinked/sie-langchain";
+ *
+ * const reranker = new SIEReranker({
+ *   baseUrl: "http://localhost:8080",
+ *   model: "jinaai/jina-reranker-v2-base-multilingual",
+ *   topK: 3,
+ * });
+ *
+ * // Rerank retrieved documents
+ * const reranked = await reranker.compressDocuments(documents, "search query");
+ *
+ * // Use in a retrieval pipeline
+ * import { ContextualCompressionRetriever } from "langchain/retrievers/contextual_compression";
+ *
+ * const compressionRetriever = new ContextualCompressionRetriever({
+ *   baseCompressor: reranker,
+ *   baseRetriever: vectorStoreRetriever,
+ * });
+ * ```
+ */
+declare class SIEReranker extends BaseDocumentCompressor {
+    private readonly model;
+    private readonly topK?;
+    private _client;
+    private readonly _ownsClient;
+    private readonly baseUrl;
+    private readonly clientOptions;
+    constructor(params?: SIERerankerParams);
+    /**
+     * Get or create the SIEClient.
+     */
+    private get client();
+    /**
+     * Rerank documents by relevance to query.
+     *
+     * @param documents - Documents to rerank.
+     * @param query - Query to rank documents against.
+     * @returns Reranked documents with relevance_score in metadata, sorted by score descending.
+     */
+    compressDocuments(documents: DocumentInterface[], query: string): Promise<DocumentInterface[]>;
+    /**
+     * Close the underlying client connection.
+     */
+    close(): Promise<void>;
+}
+
+/**
+ * SIE extraction tool for LangChain.js
+ *
+ * Provides extraction using SIE's extract endpoint:
+ * - SIEExtractor: Extraction tool implementing LangChain Tool
+ *
+ * Returns entities, relations, classifications, and detected objects.
+ *
+ * @example
+ * ```typescript
+ * import { SIEExtractor } from "@superlinked/sie-langchain";
+ *
+ * const extractor = new SIEExtractor({
+ *   baseUrl: "http://localhost:8080",
+ *   model: "urchade/gliner_multi-v2.1",
+ *   labels: ["person", "organization", "location"],
+ * });
+ *
+ * const result = await extractor.invoke("John Smith works at Acme Corp in NYC");
+ * const parsed = JSON.parse(result);
+ * console.log(parsed.entities);
+ * console.log(parsed.relations);
+ * ```
+ */
+
+/**
+ * Configuration options for SIEExtractor.
+ */
+interface SIEExtractorParams {
+    /**
+     * URL of the SIE server.
+     * @default "http://localhost:8080"
+     */
+    baseUrl?: string;
+    /**
+     * Extraction model name/ID.
+     * @default "urchade/gliner_multi-v2.1"
+     */
+    model?: string;
+    /**
+     * Optional pre-configured SIEClient instance.
+     * If provided, baseUrl and other connection options are ignored.
+     */
+    client?: SIEClient;
+    /**
+     * Labels to extract (entity types, relation types, or classification labels).
+     * @default ["person", "organization", "location"]
+     */
+    labels?: string[];
+    /**
+     * Minimum confidence threshold (0-1).
+     */
+    threshold?: number;
+    /**
+     * Target GPU type for routing (e.g., "l4", "a100-80gb").
+     */
+    gpu?: string;
+    /**
+     * Request timeout in milliseconds.
+     * @default 180000 (3 minutes)
+     */
+    timeout?: number;
+    /**
+     * Tool name for use in agents.
+     * @default "sie_extract"
+     */
+    name?: string;
+    /**
+     * Tool description for use in agents.
+     */
+    description?: string;
+}
+/**
+ * LangChain tool for extraction using SIE.
+ *
+ * Wraps SIEClient.extract() to implement the LangChain Tool interface
+ * for use in agents and chains. Returns JSON with entities, relations,
+ * classifications, and detected objects.
+ *
+ * @example
+ * ```typescript
+ * import { SIEExtractor } from "@superlinked/sie-langchain";
+ *
+ * // Direct usage
+ * const extractor = new SIEExtractor({
+ *   model: "urchade/gliner_multi-v2.1",
+ *   labels: ["person", "organization", "location"],
+ * });
+ * const result = await extractor.invoke("John Smith works at Acme Corp");
+ * const parsed = JSON.parse(result);
+ *
+ * // Use in an agent
+ * import { ChatOpenAI } from "@langchain/openai";
+ * import { createReactAgent } from "@langchain/langgraph/prebuilt";
+ *
+ * const agent = createReactAgent({
+ *   llm: new ChatOpenAI(),
+ *   tools: [extractor],
+ * });
+ * ```
+ */
+declare class SIEExtractor extends Tool {
+    name: string;
+    description: string;
+    private readonly model;
+    private readonly labels;
+    private readonly threshold?;
+    private _client;
+    private readonly _ownsClient;
+    private readonly baseUrl;
+    private readonly clientOptions;
+    constructor(params?: SIEExtractorParams);
+    /**
+     * Get or create the SIEClient.
+     */
+    private get client();
+    /**
+     * Extract structured information from text.
+     *
+     * @param text - Text to extract from.
+     * @returns JSON string with entities, relations, classifications, and objects.
+     */
+    _call(text: string): Promise<string>;
+    /**
+     * Close the underlying client connection.
+     */
+    close(): Promise<void>;
+}
 
 /**
  * SIE embeddings integration for LangChain.js
@@ -7,6 +245,8 @@ import { SIEClient, DType } from '@superlinked/sie-sdk';
  * Provides drop-in replacement for OpenAI embeddings using SIE's inference server:
  * - SIEEmbeddings: Dense embeddings for vector stores
  * - SIESparseEncoder: Sparse encoder for hybrid search
+ * - SIEReranker: Cross-encoder reranking for retrieval pipelines
+ * - SIEExtractor: Entity extraction tool for agents
  *
  * @example
  * ```typescript
@@ -118,10 +358,6 @@ declare class SIEEmbeddings extends Embeddings {
      * @returns Embedding vector as array of numbers.
      */
     embedQuery(text: string): Promise<number[]>;
-    /**
-     * Extract dense embedding from encode result.
-     */
-    private extractDense;
     /**
      * Close the underlying client connection.
      */
@@ -198,14 +434,10 @@ declare class SIESparseEncoder {
         indices: number[];
         values: number[];
     }>>;
-    /**
-     * Extract sparse embedding from encode result.
-     */
-    private extractSparse;
     /**
      * Close the underlying client connection.
      */
     close(): Promise<void>;
 }
 
-export { SIEEmbeddings, type SIEEmbeddingsParams, SIESparseEncoder, type SIESparseEncoderOptions };
+export { SIEEmbeddings, type SIEEmbeddingsParams, SIEExtractor, type SIEExtractorParams, SIEReranker, type SIERerankerParams, SIESparseEncoder, type SIESparseEncoderOptions };

package/dist/index.d.ts

@@ -1,5 +1,243 @@
 import { EmbeddingsParams, Embeddings } from '@langchain/core/embeddings';
 import { SIEClient, DType } from '@superlinked/sie-sdk';
+import { DocumentInterface } from '@langchain/core/documents';
+import { BaseDocumentCompressor } from '@langchain/core/retrievers/document_compressors';
+import { Tool } from '@langchain/core/tools';
+
+/**
+ * SIE reranker integration for LangChain.js
+ *
+ * Provides document reranking using SIE's score endpoint:
+ * - SIEReranker: Cross-encoder reranking implementing BaseDocumentCompressor
+ *
+ * @example
+ * ```typescript
+ * import { SIEReranker } from "@superlinked/sie-langchain";
+ *
+ * const reranker = new SIEReranker({
+ *   baseUrl: "http://localhost:8080",
+ *   model: "jinaai/jina-reranker-v2-base-multilingual",
+ *   topK: 3,
+ * });
+ *
+ * const reranked = await reranker.compressDocuments(documents, "search query");
+ * ```
+ */
+
+/**
+ * Configuration options for SIEReranker.
+ */
+interface SIERerankerParams {
+    /**
+     * URL of the SIE server.
+     * @default "http://localhost:8080"
+     */
+    baseUrl?: string;
+    /**
+     * Reranker model name/ID.
+     * @default "jinaai/jina-reranker-v2-base-multilingual"
+     */
+    model?: string;
+    /**
+     * Optional pre-configured SIEClient instance.
+     * If provided, baseUrl and other connection options are ignored.
+     */
+    client?: SIEClient;
+    /**
+     * Number of top documents to return. If undefined, returns all documents.
+     */
+    topK?: number;
+    /**
+     * Target GPU type for routing (e.g., "l4", "a100-80gb").
+     */
+    gpu?: string;
+    /**
+     * Request timeout in milliseconds.
+     * @default 180000 (3 minutes)
+     */
+    timeout?: number;
+}
+/**
+ * LangChain document compressor using SIE's reranking.
+ *
+ * Wraps SIEClient.score() to implement BaseDocumentCompressor.
+ *
+ * @example
+ * ```typescript
+ * import { SIEReranker } from "@superlinked/sie-langchain";
+ *
+ * const reranker = new SIEReranker({
+ *   baseUrl: "http://localhost:8080",
+ *   model: "jinaai/jina-reranker-v2-base-multilingual",
+ *   topK: 3,
+ * });
+ *
+ * // Rerank retrieved documents
+ * const reranked = await reranker.compressDocuments(documents, "search query");
+ *
+ * // Use in a retrieval pipeline
+ * import { ContextualCompressionRetriever } from "langchain/retrievers/contextual_compression";
+ *
+ * const compressionRetriever = new ContextualCompressionRetriever({
+ *   baseCompressor: reranker,
+ *   baseRetriever: vectorStoreRetriever,
+ * });
+ * ```
+ */
+declare class SIEReranker extends BaseDocumentCompressor {
+    private readonly model;
+    private readonly topK?;
+    private _client;
+    private readonly _ownsClient;
+    private readonly baseUrl;
+    private readonly clientOptions;
+    constructor(params?: SIERerankerParams);
+    /**
+     * Get or create the SIEClient.
+     */
+    private get client();
+    /**
+     * Rerank documents by relevance to query.
+     *
+     * @param documents - Documents to rerank.
+     * @param query - Query to rank documents against.
+     * @returns Reranked documents with relevance_score in metadata, sorted by score descending.
+     */
+    compressDocuments(documents: DocumentInterface[], query: string): Promise<DocumentInterface[]>;
+    /**
+     * Close the underlying client connection.
+     */
+    close(): Promise<void>;
+}
+
+/**
+ * SIE extraction tool for LangChain.js
+ *
+ * Provides extraction using SIE's extract endpoint:
+ * - SIEExtractor: Extraction tool implementing LangChain Tool
+ *
+ * Returns entities, relations, classifications, and detected objects.
+ *
+ * @example
+ * ```typescript
+ * import { SIEExtractor } from "@superlinked/sie-langchain";
+ *
+ * const extractor = new SIEExtractor({
+ *   baseUrl: "http://localhost:8080",
+ *   model: "urchade/gliner_multi-v2.1",
+ *   labels: ["person", "organization", "location"],
+ * });
+ *
+ * const result = await extractor.invoke("John Smith works at Acme Corp in NYC");
+ * const parsed = JSON.parse(result);
+ * console.log(parsed.entities);
+ * console.log(parsed.relations);
+ * ```
+ */
+
+/**
+ * Configuration options for SIEExtractor.
+ */
+interface SIEExtractorParams {
+    /**
+     * URL of the SIE server.
+     * @default "http://localhost:8080"
+     */
+    baseUrl?: string;
+    /**
+     * Extraction model name/ID.
+     * @default "urchade/gliner_multi-v2.1"
+     */
+    model?: string;
+    /**
+     * Optional pre-configured SIEClient instance.
+     * If provided, baseUrl and other connection options are ignored.
+     */
+    client?: SIEClient;
+    /**
+     * Labels to extract (entity types, relation types, or classification labels).
+     * @default ["person", "organization", "location"]
+     */
+    labels?: string[];
+    /**
+     * Minimum confidence threshold (0-1).
+     */
+    threshold?: number;
+    /**
+     * Target GPU type for routing (e.g., "l4", "a100-80gb").
+     */
+    gpu?: string;
+    /**
+     * Request timeout in milliseconds.
+     * @default 180000 (3 minutes)
+     */
+    timeout?: number;
+    /**
+     * Tool name for use in agents.
+     * @default "sie_extract"
+     */
+    name?: string;
+    /**
+     * Tool description for use in agents.
+     */
+    description?: string;
+}
+/**
+ * LangChain tool for extraction using SIE.
+ *
+ * Wraps SIEClient.extract() to implement the LangChain Tool interface
+ * for use in agents and chains. Returns JSON with entities, relations,
+ * classifications, and detected objects.
+ *
+ * @example
+ * ```typescript
+ * import { SIEExtractor } from "@superlinked/sie-langchain";
+ *
+ * // Direct usage
+ * const extractor = new SIEExtractor({
+ *   model: "urchade/gliner_multi-v2.1",
+ *   labels: ["person", "organization", "location"],
+ * });
+ * const result = await extractor.invoke("John Smith works at Acme Corp");
+ * const parsed = JSON.parse(result);
+ *
+ * // Use in an agent
+ * import { ChatOpenAI } from "@langchain/openai";
+ * import { createReactAgent } from "@langchain/langgraph/prebuilt";
+ *
+ * const agent = createReactAgent({
+ *   llm: new ChatOpenAI(),
+ *   tools: [extractor],
+ * });
+ * ```
+ */
+declare class SIEExtractor extends Tool {
+    name: string;
+    description: string;
+    private readonly model;
+    private readonly labels;
+    private readonly threshold?;
+    private _client;
+    private readonly _ownsClient;
+    private readonly baseUrl;
+    private readonly clientOptions;
+    constructor(params?: SIEExtractorParams);
+    /**
+     * Get or create the SIEClient.
+     */
+    private get client();
+    /**
+     * Extract structured information from text.
+     *
+     * @param text - Text to extract from.
+     * @returns JSON string with entities, relations, classifications, and objects.
+     */
+    _call(text: string): Promise<string>;
+    /**
+     * Close the underlying client connection.
+     */
+    close(): Promise<void>;
+}
 
 /**
  * SIE embeddings integration for LangChain.js
@@ -7,6 +245,8 @@ import { SIEClient, DType } from '@superlinked/sie-sdk';
  * Provides drop-in replacement for OpenAI embeddings using SIE's inference server:
  * - SIEEmbeddings: Dense embeddings for vector stores
  * - SIESparseEncoder: Sparse encoder for hybrid search
+ * - SIEReranker: Cross-encoder reranking for retrieval pipelines
+ * - SIEExtractor: Entity extraction tool for agents
  *
  * @example
  * ```typescript
@@ -118,10 +358,6 @@ declare class SIEEmbeddings extends Embeddings {
      * @returns Embedding vector as array of numbers.
      */
     embedQuery(text: string): Promise<number[]>;
-    /**
-     * Extract dense embedding from encode result.
-     */
-    private extractDense;
     /**
      * Close the underlying client connection.
      */
@@ -198,14 +434,10 @@ declare class SIESparseEncoder {
         indices: number[];
         values: number[];
     }>>;
-    /**
-     * Extract sparse embedding from encode result.
-     */
-    private extractSparse;
     /**
      * Close the underlying client connection.
      */
     close(): Promise<void>;
 }
 
-export { SIEEmbeddings, type SIEEmbeddingsParams, SIESparseEncoder, type SIESparseEncoderOptions };
+export { SIEEmbeddings, type SIEEmbeddingsParams, SIEExtractor, type SIEExtractorParams, SIEReranker, type SIERerankerParams, SIESparseEncoder, type SIESparseEncoderOptions };

package/dist/index.js

@@ -1,9 +1,191 @@
 // src/index.ts
 import { Embeddings } from "@langchain/core/embeddings";
 import {
-  SIEClient,
-  toNumberArray
+  SIEClient as SIEClient3,
+  denseEmbedding,
+  sparseEmbedding
 } from "@superlinked/sie-sdk";
+
+// src/rerankers.ts
+import { BaseDocumentCompressor } from "@langchain/core/retrievers/document_compressors";
+import { SIEClient } from "@superlinked/sie-sdk";
+var SIEReranker = class extends BaseDocumentCompressor {
+  model;
+  topK;
+  _client;
+  _ownsClient;
+  baseUrl;
+  clientOptions;
+  constructor(params = {}) {
+    super();
+    const {
+      baseUrl = "http://localhost:8080",
+      model = "jinaai/jina-reranker-v2-base-multilingual",
+      client,
+      topK,
+      gpu,
+      timeout = 18e4
+    } = params;
+    this.baseUrl = baseUrl;
+    this.model = model;
+    this.topK = topK;
+    this._client = client;
+    this._ownsClient = !client;
+    this.clientOptions = {
+      timeout,
+      gpu
+    };
+  }
+  /**
+   * Get or create the SIEClient.
+   */
+  get client() {
+    if (!this._client) {
+      this._client = new SIEClient(this.baseUrl, this.clientOptions);
+    }
+    return this._client;
+  }
+  /**
+   * Rerank documents by relevance to query.
+   *
+   * @param documents - Documents to rerank.
+   * @param query - Query to rank documents against.
+   * @returns Reranked documents with relevance_score in metadata, sorted by score descending.
+   */
+  async compressDocuments(documents, query) {
+    if (documents.length === 0) {
+      return [];
+    }
+    const queryItem = { text: query };
+    const docItems = documents.map((doc) => ({ text: doc.pageContent }));
+    const result = await this.client.score(this.model, queryItem, docItems);
+    const reranked = [];
+    for (const entry of result.scores) {
+      const idx = Number.parseInt(entry.itemId, 10);
+      const doc = documents[idx];
+      if (doc) {
+        reranked.push({
+          pageContent: doc.pageContent,
+          metadata: { ...doc.metadata, relevance_score: entry.score },
+          id: doc.id
+        });
+      }
+    }
+    if (this.topK !== void 0) {
+      return reranked.slice(0, this.topK);
+    }
+    return reranked;
+  }
+  /**
+   * Close the underlying client connection.
+   */
+  async close() {
+    if (this._client && this._ownsClient) {
+      await this._client.close();
+    }
+  }
+};
+
+// src/extractors.ts
+import { Tool } from "@langchain/core/tools";
+import {
+  SIEClient as SIEClient2
+} from "@superlinked/sie-sdk";
+var SIEExtractor = class extends Tool {
+  name;
+  description;
+  model;
+  labels;
+  threshold;
+  _client;
+  _ownsClient;
+  baseUrl;
+  clientOptions;
+  constructor(params = {}) {
+    const toolName = params.name ?? "sie_extract";
+    const toolDescription = params.description ?? "Extract structured information from text. Input should be text to analyze. Returns JSON with entities, relations, classifications, and detected objects.";
+    super({});
+    this.name = toolName;
+    this.description = toolDescription;
+    const {
+      baseUrl = "http://localhost:8080",
+      model = "urchade/gliner_multi-v2.1",
+      client,
+      labels = ["person", "organization", "location"],
+      threshold,
+      gpu,
+      timeout = 18e4
+    } = params;
+    this.baseUrl = baseUrl;
+    this.model = model;
+    this.labels = labels;
+    this.threshold = threshold;
+    this._client = client;
+    this._ownsClient = !client;
+    this.clientOptions = {
+      timeout,
+      gpu
+    };
+  }
+  /**
+   * Get or create the SIEClient.
+   */
+  get client() {
+    if (!this._client) {
+      this._client = new SIEClient2(this.baseUrl, this.clientOptions);
+    }
+    return this._client;
+  }
+  /**
+   * Extract structured information from text.
+   *
+   * @param text - Text to extract from.
+   * @returns JSON string with entities, relations, classifications, and objects.
+   */
+  async _call(text) {
+    const extractOptions = {
+      labels: this.labels
+    };
+    if (this.threshold !== void 0) {
+      extractOptions.threshold = this.threshold;
+    }
+    const result = await this.client.extract(this.model, { text }, extractOptions);
+    return JSON.stringify({
+      entities: result.entities.map((e) => ({
+        text: e.text,
+        label: e.label,
+        score: e.score,
+        ...e.start !== void 0 && { start: e.start },
+        ...e.end !== void 0 && { end: e.end }
+      })),
+      relations: result.relations.map((r) => ({
+        head: r.head,
+        tail: r.tail,
+        relation: r.relation,
+        score: r.score
+      })),
+      classifications: result.classifications.map((c) => ({
+        label: c.label,
+        score: c.score
+      })),
+      objects: result.objects.map((o) => ({
+        label: o.label,
+        score: o.score,
+        bbox: o.bbox
+      }))
+    });
+  }
+  /**
+   * Close the underlying client connection.
+   */
+  async close() {
+    if (this._client && this._ownsClient) {
+      await this._client.close();
+    }
+  }
+};
+
+// src/index.ts
 var SIEEmbeddings = class extends Embeddings {
   model;
   instruction;
@@ -41,7 +223,7 @@ var SIEEmbeddings = class extends Embeddings {
   get client() {
     if (!this._client) {
       const baseUrl = this.baseUrl ?? "http://localhost:8080";
-      this._client = new SIEClient(baseUrl, this.clientOptions);
+      this._client = new SIEClient3(baseUrl, this.clientOptions);
     }
     return this._client;
   }
@@ -63,7 +245,7 @@ var SIEEmbeddings = class extends Embeddings {
       isQuery: false
     };
     const results = await this.client.encode(this.model, items, options);
-    return results.map((result) => this.extractDense(result));
+    return results.map((result) => denseEmbedding(result));
   }
   /**
    * Embed a single query text.
@@ -81,17 +263,7 @@ var SIEEmbeddings = class extends Embeddings {
       isQuery: true
     };
     const result = await this.client.encode(this.model, { text }, options);
-    return this.extractDense(result);
-  }
-  /**
-   * Extract dense embedding from encode result.
-   */
-  extractDense(result) {
-    const dense = result.dense;
-    if (!dense) {
-      throw new Error("Encode result missing dense embedding");
-    }
-    return toNumberArray(dense);
+    return denseEmbedding(result);
   }
   /**
    * Close the underlying client connection.
@@ -126,7 +298,7 @@ var SIESparseEncoder = class {
    */
   get client() {
     if (!this._client) {
-      this._client = new SIEClient(this.baseUrl, this.clientOptions);
+      this._client = new SIEClient3(this.baseUrl, this.clientOptions);
     }
     return this._client;
   }
@@ -146,7 +318,7 @@ var SIESparseEncoder = class {
       isQuery: true
     };
     const results = await this.client.encode(this.model, items, options);
-    return results.map((result) => this.extractSparse(result));
+    return results.map((result) => sparseEmbedding(result));
   }
   /**
    * Encode document texts to sparse vectors.
@@ -164,20 +336,7 @@ var SIESparseEncoder = class {
       isQuery: false
     };
     const results = await this.client.encode(this.model, items, options);
-    return results.map((result) => this.extractSparse(result));
-  }
-  /**
-   * Extract sparse embedding from encode result.
-   */
-  extractSparse(result) {
-    const sparse = result.sparse;
-    if (!sparse) {
-      return { indices: [], values: [] };
-    }
-    return {
-      indices: toNumberArray(sparse.indices),
-      values: toNumberArray(sparse.values)
-    };
+    return results.map((result) => sparseEmbedding(result));
   }
   /**
    * Close the underlying client connection.
@@ -190,6 +349,8 @@ var SIESparseEncoder = class {
 };
 export {
   SIEEmbeddings,
+  SIEExtractor,
+  SIEReranker,
   SIESparseEncoder
 };
 //# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts"],"sourcesContent":["/**\n * SIE embeddings integration for LangChain.js\n *\n * Provides drop-in replacement for OpenAI embeddings using SIE's inference server:\n * - SIEEmbeddings: Dense embeddings for vector stores\n * - SIESparseEncoder: Sparse encoder for hybrid search\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * const vectors = await embeddings.embedDocuments([\"Hello world\"]);\n * const queryVector = await embeddings.embedQuery(\"What is hello?\");\n * ```\n */\n\nimport { Embeddings, type EmbeddingsParams } from \"@langchain/core/embeddings\";\nimport {\n  type DType,\n  type EncodeOptions,\n  type EncodeResult,\n  SIEClient,\n  type SIEClientOptions,\n  toNumberArray,\n} from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEEmbeddings.\n */\nexport interface SIEEmbeddingsParams extends EmbeddingsParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Optional instruction prefix for embedding (model-dependent).\n   */\n  instruction?: string;\n\n  /**\n   * Output dtype: \"float32\" (default), \"float16\", \"int8\", \"binary\".\n   */\n  outputDtype?: DType;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * LangChain Embeddings implementation using SIE.\n *\n * Wraps SIEClient.encode() to implement the LangChain Embeddings interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * // Basic usage\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * // Embed documents\n * const docVectors = await embeddings.embedDocuments([\n *   \"First document\",\n *   \"Second document\",\n * ]);\n *\n * // Embed a query (may use different encoding for asymmetric models)\n * const queryVector = await embeddings.embedQuery(\"What is the topic?\");\n *\n * // With GPU routing\n * const gpuEmbeddings = new SIEEmbeddings({\n *   baseUrl: \"https://cluster.example.com\",\n *   model: \"BAAI/bge-m3\",\n *   gpu: \"a100-80gb\",\n * });\n * ```\n */\nexport class SIEEmbeddings extends Embeddings {\n  private readonly model: string;\n  private readonly instruction?: string;\n  private readonly outputDtype?: DType;\n  private _client: SIEClient | undefined;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIEEmbeddingsParams = {}) {\n    super(params);\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      client,\n      instruction,\n      outputDtype,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.model = model;\n    this.instruction = instruction;\n    this.outputDtype = outputDtype;\n    this._client = client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n\n    // If no client provided, we'll create one lazily using baseUrl\n    if (!client) {\n      this.clientOptions.timeout = timeout;\n      this.clientOptions.gpu = gpu;\n      // Store baseUrl for lazy client creation\n      (this as { baseUrl?: string }).baseUrl = baseUrl;\n    }\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      const baseUrl = (this as { baseUrl?: string }).baseUrl ?? \"http://localhost:8080\";\n      this._client = new SIEClient(baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Embed a list of documents.\n   *\n   * @param texts - List of document texts to embed.\n   * @returns List of embedding vectors (as arrays of numbers).\n   */\n  async embedDocuments(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => this.extractDense(result));\n  }\n\n  /**\n   * Embed a single query text.\n   *\n   * For asymmetric models (like BGE-M3), this uses query-specific encoding.\n   *\n   * @param text - Query text to embed.\n   * @returns Embedding vector as array of numbers.\n   */\n  async embedQuery(text: string): Promise<number[]> {\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: true,\n    };\n\n    const result = await this.client.encode(this.model, { text }, options);\n    return this.extractDense(result as EncodeResult);\n  }\n\n  /**\n   * Extract dense embedding from encode result.\n   */\n  private extractDense(result: EncodeResult): number[] {\n    const dense = result.dense;\n    if (!dense) {\n      throw new Error(\"Encode result missing dense embedding\");\n    }\n    return toNumberArray(dense);\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n\n/**\n * Configuration options for SIESparseEncoder.\n */\nexport interface SIESparseEncoderOptions {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding. Must support sparse output.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * Sparse encoder for LangChain hybrid search.\n *\n * Compatible with PineconeHybridSearchRetriever's sparse_encoder interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings, SIESparseEncoder } from \"@superlinked/sie-langchain\";\n * import { PineconeHybridSearchRetriever } from \"@langchain/pinecone\";\n *\n * const retriever = new PineconeHybridSearchRetriever({\n *   embeddings: new SIEEmbeddings({ model: \"BAAI/bge-m3\" }),\n *   sparseEncoder: new SIESparseEncoder({ model: \"BAAI/bge-m3\" }),\n *   index: pineconeIndex,\n * });\n * ```\n */\nexport class SIESparseEncoder {\n  private readonly model: string;\n  private _client: SIEClient | undefined;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(options: SIESparseEncoderOptions = {}) {\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      gpu,\n      timeout = 180_000,\n    } = options;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Encode query texts to sparse vectors.\n   *\n   * @param texts - List of query texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeQueries(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: true,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => this.extractSparse(result));\n  }\n\n  /**\n   * Encode document texts to sparse vectors.\n   *\n   * @param texts - List of document texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeDocuments(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => this.extractSparse(result));\n  }\n\n  /**\n   * Extract sparse embedding from encode result.\n   */\n  private extractSparse(result: EncodeResult): { indices: number[]; values: number[] } {\n    const sparse = result.sparse;\n    if (!sparse) {\n      return { indices: [], values: [] };\n    }\n\n    return {\n      indices: toNumberArray(sparse.indices),\n      values: toNumberArray(sparse.values),\n    };\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n"],"mappings":";AAqBA,SAAS,kBAAyC;AAClD;AAAA,EAIE;AAAA,EAEA;AAAA,OACK;AA8EA,IAAM,gBAAN,cAA4B,WAAW;AAAA,EAC3B;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EAEjB,YAAY,SAA8B,CAAC,GAAG;AAC5C,UAAM,MAAM;AAEZ,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,QAAQ;AACb,SAAK,cAAc;AACnB,SAAK,cAAc;AACnB,SAAK,UAAU;AAEf,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAGA,QAAI,CAAC,QAAQ;AACX,WAAK,cAAc,UAAU;AAC7B,WAAK,cAAc,MAAM;AAEzB,MAAC,KAA8B,UAAU;AAAA,IAC3C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,YAAM,UAAW,KAA8B,WAAW;AAC1D,WAAK,UAAU,IAAI,UAAU,SAAS,KAAK,aAAa;AAAA,IAC1D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eAAe,OAAsC;AACzD,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,KAAK,aAAa,MAAM,CAAC;AAAA,EAC9E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,MAAM,WAAW,MAAiC;AAChD,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,SAAS,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,EAAE,KAAK,GAAG,OAAO;AACrE,WAAO,KAAK,aAAa,MAAsB;AAAA,EACjD;AAAA;AAAA;AAAA;AAAA,EAKQ,aAAa,QAAgC;AACnD,UAAM,QAAQ,OAAO;AACrB,QAAI,CAAC,OAAO;AACV,YAAM,IAAI,MAAM,uCAAuC;AAAA,IACzD;AACA,WAAO,cAAc,KAAK;AAAA,EAC5B;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;AA+CO,IAAM,mBAAN,MAAuB;AAAA,EACX;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EAEjB,YAAY,UAAmC,CAAC,GAAG;AACjD,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAI,UAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,cAAc,OAA0E;AAC5F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,KAAK,cAAc,MAAM,CAAC;AAAA,EAC/E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,gBAAgB,OAA0E;AAC9F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,KAAK,cAAc,MAAM,CAAC;AAAA,EAC/E;AAAA;AAAA;AAAA;AAAA,EAKQ,cAAc,QAA+D;AACnF,UAAM,SAAS,OAAO;AACtB,QAAI,CAAC,QAAQ;AACX,aAAO,EAAE,SAAS,CAAC,GAAG,QAAQ,CAAC,EAAE;AAAA,IACnC;AAEA,WAAO;AAAA,MACL,SAAS,cAAc,OAAO,OAAO;AAAA,MACrC,QAAQ,cAAc,OAAO,MAAM;AAAA,IACrC;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;","names":[]}
+{"version":3,"sources":["../src/index.ts","../src/rerankers.ts","../src/extractors.ts"],"sourcesContent":["/**\n * SIE embeddings integration for LangChain.js\n *\n * Provides drop-in replacement for OpenAI embeddings using SIE's inference server:\n * - SIEEmbeddings: Dense embeddings for vector stores\n * - SIESparseEncoder: Sparse encoder for hybrid search\n * - SIEReranker: Cross-encoder reranking for retrieval pipelines\n * - SIEExtractor: Entity extraction tool for agents\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * const vectors = await embeddings.embedDocuments([\"Hello world\"]);\n * const queryVector = await embeddings.embedQuery(\"What is hello?\");\n * ```\n */\n\nimport { Embeddings, type EmbeddingsParams } from \"@langchain/core/embeddings\";\nimport {\n  type DType,\n  type EncodeOptions,\n  type EncodeResult,\n  SIEClient,\n  type SIEClientOptions,\n  denseEmbedding,\n  sparseEmbedding,\n} from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEEmbeddings.\n */\nexport interface SIEEmbeddingsParams extends EmbeddingsParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Optional instruction prefix for embedding (model-dependent).\n   */\n  instruction?: string;\n\n  /**\n   * Output dtype: \"float32\" (default), \"float16\", \"int8\", \"binary\".\n   */\n  outputDtype?: DType;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * LangChain Embeddings implementation using SIE.\n *\n * Wraps SIEClient.encode() to implement the LangChain Embeddings interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings } from \"@superlinked/sie-langchain\";\n *\n * // Basic usage\n * const embeddings = new SIEEmbeddings({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"BAAI/bge-m3\",\n * });\n *\n * // Embed documents\n * const docVectors = await embeddings.embedDocuments([\n *   \"First document\",\n *   \"Second document\",\n * ]);\n *\n * // Embed a query (may use different encoding for asymmetric models)\n * const queryVector = await embeddings.embedQuery(\"What is the topic?\");\n *\n * // With GPU routing\n * const gpuEmbeddings = new SIEEmbeddings({\n *   baseUrl: \"https://cluster.example.com\",\n *   model: \"BAAI/bge-m3\",\n *   gpu: \"a100-80gb\",\n * });\n * ```\n */\nexport class SIEEmbeddings extends Embeddings {\n  private readonly model: string;\n  private readonly instruction?: string;\n  private readonly outputDtype?: DType;\n  private _client: SIEClient | undefined;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIEEmbeddingsParams = {}) {\n    super(params);\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      client,\n      instruction,\n      outputDtype,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.model = model;\n    this.instruction = instruction;\n    this.outputDtype = outputDtype;\n    this._client = client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n\n    // If no client provided, we'll create one lazily using baseUrl\n    if (!client) {\n      this.clientOptions.timeout = timeout;\n      this.clientOptions.gpu = gpu;\n      // Store baseUrl for lazy client creation\n      (this as { baseUrl?: string }).baseUrl = baseUrl;\n    }\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      const baseUrl = (this as { baseUrl?: string }).baseUrl ?? \"http://localhost:8080\";\n      this._client = new SIEClient(baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Embed a list of documents.\n   *\n   * @param texts - List of document texts to embed.\n   * @returns List of embedding vectors (as arrays of numbers).\n   */\n  async embedDocuments(texts: string[]): Promise<number[][]> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => denseEmbedding(result));\n  }\n\n  /**\n   * Embed a single query text.\n   *\n   * For asymmetric models (like BGE-M3), this uses query-specific encoding.\n   *\n   * @param text - Query text to embed.\n   * @returns Embedding vector as array of numbers.\n   */\n  async embedQuery(text: string): Promise<number[]> {\n    const options: EncodeOptions = {\n      outputTypes: [\"dense\"],\n      instruction: this.instruction,\n      outputDtype: this.outputDtype,\n      isQuery: true,\n    };\n\n    const result = await this.client.encode(this.model, { text }, options);\n    return denseEmbedding(result as EncodeResult);\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n\n/**\n * Configuration options for SIESparseEncoder.\n */\nexport interface SIESparseEncoderOptions {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Model name/ID to use for encoding. Must support sparse output.\n   * @default \"BAAI/bge-m3\"\n   */\n  model?: string;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * Sparse encoder for LangChain hybrid search.\n *\n * Compatible with PineconeHybridSearchRetriever's sparse_encoder interface.\n *\n * @example\n * ```typescript\n * import { SIEEmbeddings, SIESparseEncoder } from \"@superlinked/sie-langchain\";\n * import { PineconeHybridSearchRetriever } from \"@langchain/pinecone\";\n *\n * const retriever = new PineconeHybridSearchRetriever({\n *   embeddings: new SIEEmbeddings({ model: \"BAAI/bge-m3\" }),\n *   sparseEncoder: new SIESparseEncoder({ model: \"BAAI/bge-m3\" }),\n *   index: pineconeIndex,\n * });\n * ```\n */\nexport class SIESparseEncoder {\n  private readonly model: string;\n  private _client: SIEClient | undefined;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(options: SIESparseEncoderOptions = {}) {\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"BAAI/bge-m3\",\n      gpu,\n      timeout = 180_000,\n    } = options;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Encode query texts to sparse vectors.\n   *\n   * @param texts - List of query texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeQueries(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: true,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => sparseEmbedding(result));\n  }\n\n  /**\n   * Encode document texts to sparse vectors.\n   *\n   * @param texts - List of document texts to encode.\n   * @returns List of objects with \"indices\" and \"values\" arrays.\n   */\n  async encodeDocuments(texts: string[]): Promise<Array<{ indices: number[]; values: number[] }>> {\n    if (texts.length === 0) {\n      return [];\n    }\n\n    const items = texts.map((text) => ({ text }));\n    const options: EncodeOptions = {\n      outputTypes: [\"sparse\"],\n      isQuery: false,\n    };\n\n    const results = await this.client.encode(this.model, items, options);\n    return (results as EncodeResult[]).map((result) => sparseEmbedding(result));\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client) {\n      await this._client.close();\n    }\n  }\n}\n\nexport { SIEReranker, type SIERerankerParams } from \"./rerankers.js\";\nexport { SIEExtractor, type SIEExtractorParams } from \"./extractors.js\";\n","/**\n * SIE reranker integration for LangChain.js\n *\n * Provides document reranking using SIE's score endpoint:\n * - SIEReranker: Cross-encoder reranking implementing BaseDocumentCompressor\n *\n * @example\n * ```typescript\n * import { SIEReranker } from \"@superlinked/sie-langchain\";\n *\n * const reranker = new SIEReranker({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"jinaai/jina-reranker-v2-base-multilingual\",\n *   topK: 3,\n * });\n *\n * const reranked = await reranker.compressDocuments(documents, \"search query\");\n * ```\n */\n\nimport type { DocumentInterface } from \"@langchain/core/documents\";\nimport { BaseDocumentCompressor } from \"@langchain/core/retrievers/document_compressors\";\nimport { SIEClient, type SIEClientOptions } from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEReranker.\n */\nexport interface SIERerankerParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Reranker model name/ID.\n   * @default \"jinaai/jina-reranker-v2-base-multilingual\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Number of top documents to return. If undefined, returns all documents.\n   */\n  topK?: number;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n}\n\n/**\n * LangChain document compressor using SIE's reranking.\n *\n * Wraps SIEClient.score() to implement BaseDocumentCompressor.\n *\n * @example\n * ```typescript\n * import { SIEReranker } from \"@superlinked/sie-langchain\";\n *\n * const reranker = new SIEReranker({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"jinaai/jina-reranker-v2-base-multilingual\",\n *   topK: 3,\n * });\n *\n * // Rerank retrieved documents\n * const reranked = await reranker.compressDocuments(documents, \"search query\");\n *\n * // Use in a retrieval pipeline\n * import { ContextualCompressionRetriever } from \"langchain/retrievers/contextual_compression\";\n *\n * const compressionRetriever = new ContextualCompressionRetriever({\n *   baseCompressor: reranker,\n *   baseRetriever: vectorStoreRetriever,\n * });\n * ```\n */\nexport class SIEReranker extends BaseDocumentCompressor {\n  private readonly model: string;\n  private readonly topK?: number;\n  private _client: SIEClient | undefined;\n  private readonly _ownsClient: boolean;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIERerankerParams = {}) {\n    super();\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"jinaai/jina-reranker-v2-base-multilingual\",\n      client,\n      topK,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.topK = topK;\n    this._client = client;\n    this._ownsClient = !client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Rerank documents by relevance to query.\n   *\n   * @param documents - Documents to rerank.\n   * @param query - Query to rank documents against.\n   * @returns Reranked documents with relevance_score in metadata, sorted by score descending.\n   */\n  async compressDocuments(\n    documents: DocumentInterface[],\n    query: string,\n  ): Promise<DocumentInterface[]> {\n    if (documents.length === 0) {\n      return [];\n    }\n\n    const queryItem = { text: query };\n    const docItems = documents.map((doc) => ({ text: doc.pageContent }));\n\n    const result = await this.client.score(this.model, queryItem, docItems);\n\n    // Map score entries back to documents with relevance_score in metadata.\n    // ScoreResult.scores are already sorted by score descending.\n    const reranked: DocumentInterface[] = [];\n    for (const entry of result.scores) {\n      const idx = Number.parseInt(entry.itemId, 10);\n      const doc = documents[idx];\n      if (doc) {\n        reranked.push({\n          pageContent: doc.pageContent,\n          metadata: { ...doc.metadata, relevance_score: entry.score },\n          id: doc.id,\n        });\n      }\n    }\n\n    if (this.topK !== undefined) {\n      return reranked.slice(0, this.topK);\n    }\n    return reranked;\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client && this._ownsClient) {\n      await this._client.close();\n    }\n  }\n}\n","/**\n * SIE extraction tool for LangChain.js\n *\n * Provides extraction using SIE's extract endpoint:\n * - SIEExtractor: Extraction tool implementing LangChain Tool\n *\n * Returns entities, relations, classifications, and detected objects.\n *\n * @example\n * ```typescript\n * import { SIEExtractor } from \"@superlinked/sie-langchain\";\n *\n * const extractor = new SIEExtractor({\n *   baseUrl: \"http://localhost:8080\",\n *   model: \"urchade/gliner_multi-v2.1\",\n *   labels: [\"person\", \"organization\", \"location\"],\n * });\n *\n * const result = await extractor.invoke(\"John Smith works at Acme Corp in NYC\");\n * const parsed = JSON.parse(result);\n * console.log(parsed.entities);\n * console.log(parsed.relations);\n * ```\n */\n\nimport { Tool } from \"@langchain/core/tools\";\nimport {\n  type ExtractOptions,\n  type ExtractResult,\n  SIEClient,\n  type SIEClientOptions,\n} from \"@superlinked/sie-sdk\";\n\n/**\n * Configuration options for SIEExtractor.\n */\nexport interface SIEExtractorParams {\n  /**\n   * URL of the SIE server.\n   * @default \"http://localhost:8080\"\n   */\n  baseUrl?: string;\n\n  /**\n   * Extraction model name/ID.\n   * @default \"urchade/gliner_multi-v2.1\"\n   */\n  model?: string;\n\n  /**\n   * Optional pre-configured SIEClient instance.\n   * If provided, baseUrl and other connection options are ignored.\n   */\n  client?: SIEClient;\n\n  /**\n   * Labels to extract (entity types, relation types, or classification labels).\n   * @default [\"person\", \"organization\", \"location\"]\n   */\n  labels?: string[];\n\n  /**\n   * Minimum confidence threshold (0-1).\n   */\n  threshold?: number;\n\n  /**\n   * Target GPU type for routing (e.g., \"l4\", \"a100-80gb\").\n   */\n  gpu?: string;\n\n  /**\n   * Request timeout in milliseconds.\n   * @default 180000 (3 minutes)\n   */\n  timeout?: number;\n\n  /**\n   * Tool name for use in agents.\n   * @default \"sie_extract\"\n   */\n  name?: string;\n\n  /**\n   * Tool description for use in agents.\n   */\n  description?: string;\n}\n\n/**\n * LangChain tool for extraction using SIE.\n *\n * Wraps SIEClient.extract() to implement the LangChain Tool interface\n * for use in agents and chains. Returns JSON with entities, relations,\n * classifications, and detected objects.\n *\n * @example\n * ```typescript\n * import { SIEExtractor } from \"@superlinked/sie-langchain\";\n *\n * // Direct usage\n * const extractor = new SIEExtractor({\n *   model: \"urchade/gliner_multi-v2.1\",\n *   labels: [\"person\", \"organization\", \"location\"],\n * });\n * const result = await extractor.invoke(\"John Smith works at Acme Corp\");\n * const parsed = JSON.parse(result);\n *\n * // Use in an agent\n * import { ChatOpenAI } from \"@langchain/openai\";\n * import { createReactAgent } from \"@langchain/langgraph/prebuilt\";\n *\n * const agent = createReactAgent({\n *   llm: new ChatOpenAI(),\n *   tools: [extractor],\n * });\n * ```\n */\nexport class SIEExtractor extends Tool {\n  name: string;\n  description: string;\n\n  private readonly model: string;\n  private readonly labels: string[];\n  private readonly threshold?: number;\n  private _client: SIEClient | undefined;\n  private readonly _ownsClient: boolean;\n  private readonly baseUrl: string;\n  private readonly clientOptions: SIEClientOptions;\n\n  constructor(params: SIEExtractorParams = {}) {\n    const toolName = params.name ?? \"sie_extract\";\n    const toolDescription =\n      params.description ??\n      \"Extract structured information from text. \" +\n        \"Input should be text to analyze. \" +\n        \"Returns JSON with entities, relations, classifications, and detected objects.\";\n\n    super({});\n\n    this.name = toolName;\n    this.description = toolDescription;\n\n    const {\n      baseUrl = \"http://localhost:8080\",\n      model = \"urchade/gliner_multi-v2.1\",\n      client,\n      labels = [\"person\", \"organization\", \"location\"],\n      threshold,\n      gpu,\n      timeout = 180_000,\n    } = params;\n\n    this.baseUrl = baseUrl;\n    this.model = model;\n    this.labels = labels;\n    this.threshold = threshold;\n    this._client = client;\n    this._ownsClient = !client;\n\n    this.clientOptions = {\n      timeout,\n      gpu,\n    };\n  }\n\n  /**\n   * Get or create the SIEClient.\n   */\n  private get client(): SIEClient {\n    if (!this._client) {\n      this._client = new SIEClient(this.baseUrl, this.clientOptions);\n    }\n    return this._client;\n  }\n\n  /**\n   * Extract structured information from text.\n   *\n   * @param text - Text to extract from.\n   * @returns JSON string with entities, relations, classifications, and objects.\n   */\n  async _call(text: string): Promise<string> {\n    const extractOptions: ExtractOptions = {\n      labels: this.labels,\n    };\n    if (this.threshold !== undefined) {\n      extractOptions.threshold = this.threshold;\n    }\n\n    const result: ExtractResult = await this.client.extract(this.model, { text }, extractOptions);\n\n    return JSON.stringify({\n      entities: result.entities.map((e) => ({\n        text: e.text,\n        label: e.label,\n        score: e.score,\n        ...(e.start !== undefined && { start: e.start }),\n        ...(e.end !== undefined && { end: e.end }),\n      })),\n      relations: result.relations.map((r) => ({\n        head: r.head,\n        tail: r.tail,\n        relation: r.relation,\n        score: r.score,\n      })),\n      classifications: result.classifications.map((c) => ({\n        label: c.label,\n        score: c.score,\n      })),\n      objects: result.objects.map((o) => ({\n        label: o.label,\n        score: o.score,\n        bbox: o.bbox,\n      })),\n    });\n  }\n\n  /**\n   * Close the underlying client connection.\n   */\n  async close(): Promise<void> {\n    if (this._client && this._ownsClient) {\n      await this._client.close();\n    }\n  }\n}\n"],"mappings":";AAuBA,SAAS,kBAAyC;AAClD;AAAA,EAIE,aAAAA;AAAA,EAEA;AAAA,EACA;AAAA,OACK;;;ACXP,SAAS,8BAA8B;AACvC,SAAS,iBAAwC;AAoE1C,IAAM,cAAN,cAA0B,uBAAuB;AAAA,EACrC;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EACA;AAAA,EAEjB,YAAY,SAA4B,CAAC,GAAG;AAC1C,UAAM;AAEN,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,OAAO;AACZ,SAAK,UAAU;AACf,SAAK,cAAc,CAAC;AAEpB,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAI,UAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,kBACJ,WACA,OAC8B;AAC9B,QAAI,UAAU,WAAW,GAAG;AAC1B,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,YAAY,EAAE,MAAM,MAAM;AAChC,UAAM,WAAW,UAAU,IAAI,CAAC,SAAS,EAAE,MAAM,IAAI,YAAY,EAAE;AAEnE,UAAM,SAAS,MAAM,KAAK,OAAO,MAAM,KAAK,OAAO,WAAW,QAAQ;AAItE,UAAM,WAAgC,CAAC;AACvC,eAAW,SAAS,OAAO,QAAQ;AACjC,YAAM,MAAM,OAAO,SAAS,MAAM,QAAQ,EAAE;AAC5C,YAAM,MAAM,UAAU,GAAG;AACzB,UAAI,KAAK;AACP,iBAAS,KAAK;AAAA,UACZ,aAAa,IAAI;AAAA,UACjB,UAAU,EAAE,GAAG,IAAI,UAAU,iBAAiB,MAAM,MAAM;AAAA,UAC1D,IAAI,IAAI;AAAA,QACV,CAAC;AAAA,MACH;AAAA,IACF;AAEA,QAAI,KAAK,SAAS,QAAW;AAC3B,aAAO,SAAS,MAAM,GAAG,KAAK,IAAI;AAAA,IACpC;AACA,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,WAAW,KAAK,aAAa;AACpC,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;;;AC5JA,SAAS,YAAY;AACrB;AAAA,EAGE,aAAAC;AAAA,OAEK;AAuFA,IAAM,eAAN,cAA2B,KAAK;AAAA,EACrC;AAAA,EACA;AAAA,EAEiB;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EACA;AAAA,EAEjB,YAAY,SAA6B,CAAC,GAAG;AAC3C,UAAM,WAAW,OAAO,QAAQ;AAChC,UAAM,kBACJ,OAAO,eACP;AAIF,UAAM,CAAC,CAAC;AAER,SAAK,OAAO;AACZ,SAAK,cAAc;AAEnB,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA,SAAS,CAAC,UAAU,gBAAgB,UAAU;AAAA,MAC9C;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,SAAS;AACd,SAAK,YAAY;AACjB,SAAK,UAAU;AACf,SAAK,cAAc,CAAC;AAEpB,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAIA,WAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,MAAM,MAA+B;AACzC,UAAM,iBAAiC;AAAA,MACrC,QAAQ,KAAK;AAAA,IACf;AACA,QAAI,KAAK,cAAc,QAAW;AAChC,qBAAe,YAAY,KAAK;AAAA,IAClC;AAEA,UAAM,SAAwB,MAAM,KAAK,OAAO,QAAQ,KAAK,OAAO,EAAE,KAAK,GAAG,cAAc;AAE5F,WAAO,KAAK,UAAU;AAAA,MACpB,UAAU,OAAO,SAAS,IAAI,CAAC,OAAO;AAAA,QACpC,MAAM,EAAE;AAAA,QACR,OAAO,EAAE;AAAA,QACT,OAAO,EAAE;AAAA,QACT,GAAI,EAAE,UAAU,UAAa,EAAE,OAAO,EAAE,MAAM;AAAA,QAC9C,GAAI,EAAE,QAAQ,UAAa,EAAE,KAAK,EAAE,IAAI;AAAA,MAC1C,EAAE;AAAA,MACF,WAAW,OAAO,UAAU,IAAI,CAAC,OAAO;AAAA,QACtC,MAAM,EAAE;AAAA,QACR,MAAM,EAAE;AAAA,QACR,UAAU,EAAE;AAAA,QACZ,OAAO,EAAE;AAAA,MACX,EAAE;AAAA,MACF,iBAAiB,OAAO,gBAAgB,IAAI,CAAC,OAAO;AAAA,QAClD,OAAO,EAAE;AAAA,QACT,OAAO,EAAE;AAAA,MACX,EAAE;AAAA,MACF,SAAS,OAAO,QAAQ,IAAI,CAAC,OAAO;AAAA,QAClC,OAAO,EAAE;AAAA,QACT,OAAO,EAAE;AAAA,QACT,MAAM,EAAE;AAAA,MACV,EAAE;AAAA,IACJ,CAAC;AAAA,EACH;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,WAAW,KAAK,aAAa;AACpC,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;;;AFpHO,IAAM,gBAAN,cAA4B,WAAW;AAAA,EAC3B;AAAA,EACA;AAAA,EACA;AAAA,EACT;AAAA,EACS;AAAA,EAEjB,YAAY,SAA8B,CAAC,GAAG;AAC5C,UAAM,MAAM;AAEZ,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,QAAQ;AACb,SAAK,cAAc;AACnB,SAAK,cAAc;AACnB,SAAK,UAAU;AAEf,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAGA,QAAI,CAAC,QAAQ;AACX,WAAK,cAAc,UAAU;AAC7B,WAAK,cAAc,MAAM;AAEzB,MAAC,KAA8B,UAAU;AAAA,IAC3C;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,YAAM,UAAW,KAA8B,WAAW;AAC1D,WAAK,UAAU,IAAIC,WAAU,SAAS,KAAK,aAAa;AAAA,IAC1D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eAAe,OAAsC;AACzD,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,eAAe,MAAM,CAAC;AAAA,EAC3E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA,MAAM,WAAW,MAAiC;AAChD,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,OAAO;AAAA,MACrB,aAAa,KAAK;AAAA,MAClB,aAAa,KAAK;AAAA,MAClB,SAAS;AAAA,IACX;AAEA,UAAM,SAAS,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,EAAE,KAAK,GAAG,OAAO;AACrE,WAAO,eAAe,MAAsB;AAAA,EAC9C;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;AA+CO,IAAM,mBAAN,MAAuB;AAAA,EACX;AAAA,EACT;AAAA,EACS;AAAA,EACA;AAAA,EAEjB,YAAY,UAAmC,CAAC,GAAG;AACjD,UAAM;AAAA,MACJ,UAAU;AAAA,MACV,QAAQ;AAAA,MACR;AAAA,MACA,UAAU;AAAA,IACZ,IAAI;AAEJ,SAAK,UAAU;AACf,SAAK,QAAQ;AACb,SAAK,gBAAgB;AAAA,MACnB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,IAAY,SAAoB;AAC9B,QAAI,CAAC,KAAK,SAAS;AACjB,WAAK,UAAU,IAAIA,WAAU,KAAK,SAAS,KAAK,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,cAAc,OAA0E;AAC5F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,gBAAgB,MAAM,CAAC;AAAA,EAC5E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,gBAAgB,OAA0E;AAC9F,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,CAAC;AAAA,IACV;AAEA,UAAM,QAAQ,MAAM,IAAI,CAAC,UAAU,EAAE,KAAK,EAAE;AAC5C,UAAM,UAAyB;AAAA,MAC7B,aAAa,CAAC,QAAQ;AAAA,MACtB,SAAS;AAAA,IACX;AAEA,UAAM,UAAU,MAAM,KAAK,OAAO,OAAO,KAAK,OAAO,OAAO,OAAO;AACnE,WAAQ,QAA2B,IAAI,CAAC,WAAW,gBAAgB,MAAM,CAAC;AAAA,EAC5E;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,QAAuB;AAC3B,QAAI,KAAK,SAAS;AAChB,YAAM,KAAK,QAAQ,MAAM;AAAA,IAC3B;AAAA,EACF;AACF;","names":["SIEClient","SIEClient","SIEClient"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@superlinked/sie-langchain",
-  "version": "0.1.9",
+  "version": "0.1.10",
   "description": "SIE embeddings integration for LangChain.js",
   "type": "module",
   "main": "./dist/index.cjs",
@@ -17,7 +17,7 @@
     "dist"
   ],
   "dependencies": {
-    "@superlinked/sie-sdk": "0.1.9"
+    "@superlinked/sie-sdk": "0.1.10"
   },
   "peerDependencies": {
     "@langchain/core": ">=0.2.0"