@langchain/classic 1.0.17 → 1.0.18

package/dist/retrievers/matryoshka_retriever.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"matryoshka_retriever.cjs","names":["VectorStoreRetriever","fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>","embeddedQuery: number[]","smallResults: DocumentInterface[]","largeEmbeddings: Array<number[]>","func: () => Array<number[]>","query: string","documents: DocumentInterface[]","options?: AddDocumentOptions","newDocuments: Array<DocumentInterface>"],"sources":["../../src/retrievers/matryoshka_retriever.ts"],"sourcesContent":["import { DocumentInterface } from \"@langchain/core/documents\";\nimport { Embeddings } from \"@langchain/core/embeddings\";\nimport {\n  cosineSimilarity,\n  euclideanDistance,\n  innerProduct,\n} from \"@langchain/core/utils/math\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\n\n/**\n * Type for options when adding a document to the VectorStore.\n */\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\ntype AddDocumentOptions = Record<string, any>;\n\nexport interface MatryoshkaRetrieverFields {\n  /**\n   * The number of documents to retrieve from the small store.\n   * @default 50\n   */\n  smallK?: number;\n  /**\n   * The number of documents to retrieve from the large store.\n   * @default 8\n   */\n  largeK?: number;\n  /**\n   * The metadata key to store the larger embeddings.\n   * @default \"lc_large_embedding\"\n   */\n  largeEmbeddingKey?: string;\n  /**\n   * The embedding model to use when generating the large\n   * embeddings.\n   */\n  largeEmbeddingModel: Embeddings;\n  /**\n   * The type of search to perform using the large embeddings.\n   * @default \"cosine\"\n   */\n  searchType?: \"cosine\" | \"innerProduct\" | \"euclidean\";\n}\n\n/**\n * A retriever that uses two sets of embeddings to perform adaptive retrieval. Based\n * off of the \"Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval\"\n * blog post {@link https://supabase.com/blog/matryoshka-embeddings}.\n *\n *\n * This class performs \"Adaptive Retrieval\" for searching text embeddings efficiently using the\n * Matryoshka Representation Learning (MRL) technique. It retrieves documents similar to a query\n * embedding in two steps:\n *\n * First-pass: Uses a lower dimensional sub-vector from the MRL embedding for an initial, fast,\n * but less accurate search.\n *\n * Second-pass: Re-ranks the top results from the first pass using the full, high-dimensional\n * embedding for higher accuracy.\n *\n *\n * This code implements MRL embeddings for efficient vector search by combining faster,\n * lower-dimensional initial search with accurate, high-dimensional re-ranking.\n */\nexport class MatryoshkaRetriever<\n  Store extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<Store> {\n  smallK = 50;\n\n  largeK = 8;\n\n  largeEmbeddingKey = \"lc_large_embedding\";\n\n  largeEmbeddingModel: Embeddings;\n\n  searchType: \"cosine\" | \"innerProduct\" | \"euclidean\" = \"cosine\";\n\n  constructor(\n    fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>\n  ) {\n    super(fields);\n    this.smallK = fields.smallK ?? this.smallK;\n    this.largeK = fields.largeK ?? this.largeK;\n    this.largeEmbeddingKey = fields.largeEmbeddingKey ?? this.largeEmbeddingKey;\n    this.largeEmbeddingModel = fields.largeEmbeddingModel;\n    this.searchType = fields.searchType ?? this.searchType;\n  }\n\n  /**\n   * Ranks documents based on their similarity to a query embedding using larger embeddings.\n   *\n   * This method takes a query embedding and a list of documents (smallResults) as input. Each document\n   * in the smallResults array has previously been associated with a large embedding stored in its metadata.\n   * Depending on the `searchType` (cosine, innerProduct, or euclidean), it calculates the similarity scores\n   * between the query embedding and each document's large embedding. It then ranks the documents based on\n   * these similarity scores, from the most similar to the least similar.\n   *\n   * The method returns a promise that resolves to an array of the top `largeK` documents, where `largeK`\n   * is a class property defining the number of documents to return. This subset of documents is determined\n   * by sorting the entire list of documents based on their similarity scores and then selecting the top\n   * `largeK` documents.\n   *\n   * @param {number[]} embeddedQuery The embedding of the query, represented as an array of numbers.\n   * @param {DocumentInterface[]} smallResults An array of documents, each with metadata that includes a large embedding for similarity comparison.\n   * @returns {Promise<DocumentInterface[]>} A promise that resolves to an array of the top `largeK` ranked documents based on their similarity to the query embedding.\n   */\n  private _rankByLargeEmbeddings(\n    embeddedQuery: number[],\n    smallResults: DocumentInterface[]\n  ): DocumentInterface[] {\n    const largeEmbeddings: Array<number[]> = smallResults.map((doc) =>\n      JSON.parse(doc.metadata[this.largeEmbeddingKey])\n    );\n    let func: () => Array<number[]>;\n\n    switch (this.searchType) {\n      case \"cosine\":\n        func = () => cosineSimilarity([embeddedQuery], largeEmbeddings);\n        break;\n      case \"innerProduct\":\n        func = () => innerProduct([embeddedQuery], largeEmbeddings);\n        break;\n      case \"euclidean\":\n        func = () => euclideanDistance([embeddedQuery], largeEmbeddings);\n        break;\n      default:\n        throw new Error(`Unknown search type: ${this.searchType}`);\n    }\n\n    // Calculate the similarity scores between the query embedding and the large embeddings\n    const [similarityScores] = func();\n\n    // Create an array of indices from 0 to N-1, where N is the number of documents\n    let indices = Array.from(\n      { length: smallResults.length },\n      (_, index) => index\n    );\n\n    indices = indices\n      .map((v, i) => [similarityScores[i], v])\n      .sort(([a], [b]) => b - a)\n      .slice(0, this.largeK)\n      .map(([, i]) => i);\n\n    return indices.map((i) => smallResults[i]);\n  }\n\n  async _getRelevantDocuments(query: string): Promise<DocumentInterface[]> {\n    const [embeddedQuery, smallResults] = await Promise.all([\n      this.largeEmbeddingModel.embedQuery(query),\n      this.vectorStore.similaritySearch(query, this.smallK, this.filter),\n    ]);\n\n    return this._rankByLargeEmbeddings(embeddedQuery, smallResults);\n  }\n\n  /**\n   * Override the default `addDocuments` method to embed the documents twice,\n   * once using the larger embeddings model, and then again using the default\n   * embedding model linked to the vector store.\n   *\n   * @param {DocumentInterface[]} documents - An array of documents to add to the vector store.\n   * @param {AddDocumentOptions} options - An optional object containing additional options for adding documents.\n   * @returns {Promise<string[] | void>} A promise that resolves to an array of the document IDs that were added to the vector store.\n   */\n  override addDocuments = async (\n    documents: DocumentInterface[],\n    options?: AddDocumentOptions\n  ): Promise<string[] | void> => {\n    // Insure documents metadata does not contain the large embedding key\n    if (documents.some((doc) => this.largeEmbeddingKey in doc.metadata)) {\n      throw new Error(\n        `All documents must not contain the large embedding key: ${this.largeEmbeddingKey} in their metadata.`\n      );\n    }\n\n    const allDocPageContent = documents.map((doc) => doc.pageContent);\n    const allDocLargeEmbeddings =\n      await this.largeEmbeddingModel.embedDocuments(allDocPageContent);\n\n    const newDocuments: Array<DocumentInterface> = documents.map(\n      (doc, idx) => ({\n        ...doc,\n        metadata: {\n          ...doc.metadata,\n          [this.largeEmbeddingKey]: JSON.stringify(allDocLargeEmbeddings[idx]),\n        },\n      })\n    );\n\n    return this.vectorStore.addDocuments(newDocuments, options);\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAmEA,IAAa,sBAAb,cAEUA,mDAA4B;CACpC,SAAS;CAET,SAAS;CAET,oBAAoB;CAEpB;CAEA,aAAsD;CAEtD,YACEC,QACA;EACA,MAAM,OAAO;EACb,KAAK,SAAS,OAAO,UAAU,KAAK;EACpC,KAAK,SAAS,OAAO,UAAU,KAAK;EACpC,KAAK,oBAAoB,OAAO,qBAAqB,KAAK;EAC1D,KAAK,sBAAsB,OAAO;EAClC,KAAK,aAAa,OAAO,cAAc,KAAK;CAC7C;;;;;;;;;;;;;;;;;;;CAoBD,AAAQ,uBACNC,eACAC,cACqB;EACrB,MAAMC,kBAAmC,aAAa,IAAI,CAAC,QACzD,KAAK,MAAM,IAAI,SAAS,KAAK,mBAAmB,CACjD;EACD,IAAIC;AAEJ,UAAQ,KAAK,YAAb;GACE,KAAK;IACH,OAAO,wDAAuB,CAAC,aAAc,GAAE,gBAAgB;AAC/D;GACF,KAAK;IACH,OAAO,oDAAmB,CAAC,aAAc,GAAE,gBAAgB;AAC3D;GACF,KAAK;IACH,OAAO,yDAAwB,CAAC,aAAc,GAAE,gBAAgB;AAChE;GACF,QACE,OAAM,IAAI,MAAM,CAAC,qBAAqB,EAAE,KAAK,YAAY;EAC5D;EAGD,MAAM,CAAC,iBAAiB,GAAG,MAAM;EAGjC,IAAI,UAAU,MAAM,KAClB,EAAE,QAAQ,aAAa,OAAQ,GAC/B,CAAC,GAAG,UAAU,MACf;EAED,UAAU,QACP,IAAI,CAAC,GAAG,MAAM,CAAC,iBAAiB,IAAI,CAAE,EAAC,CACvC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,KAAK,IAAI,EAAE,CACzB,MAAM,GAAG,KAAK,OAAO,CACrB,IAAI,CAAC,GAAG,EAAE,KAAK,EAAE;AAEpB,SAAO,QAAQ,IAAI,CAAC,MAAM,aAAa,GAAG;CAC3C;CAED,MAAM,sBAAsBC,OAA6C;EACvE,MAAM,CAAC,eAAe,aAAa,GAAG,MAAM,QAAQ,IAAI,CACtD,KAAK,oBAAoB,WAAW,MAAM,EAC1C,KAAK,YAAY,iBAAiB,OAAO,KAAK,QAAQ,KAAK,OAAO,AACnE,EAAC;AAEF,SAAO,KAAK,uBAAuB,eAAe,aAAa;CAChE;;;;;;;;;;CAWD,AAAS,eAAe,OACtBC,WACAC,YAC6B;AAE7B,MAAI,UAAU,KAAK,CAAC,QAAQ,KAAK,qBAAqB,IAAI,SAAS,CACjE,OAAM,IAAI,MACR,CAAC,wDAAwD,EAAE,KAAK,kBAAkB,mBAAmB,CAAC;EAI1G,MAAM,oBAAoB,UAAU,IAAI,CAAC,QAAQ,IAAI,YAAY;EACjE,MAAM,wBACJ,MAAM,KAAK,oBAAoB,eAAe,kBAAkB;EAElE,MAAMC,eAAyC,UAAU,IACvD,CAAC,KAAK,SAAS;GACb,GAAG;GACH,UAAU;IACR,GAAG,IAAI;KACN,KAAK,oBAAoB,KAAK,UAAU,sBAAsB,KAAK;GACrE;EACF,GACF;AAED,SAAO,KAAK,YAAY,aAAa,cAAc,QAAQ;CAC5D;AACF"}
+{"version":3,"file":"matryoshka_retriever.cjs","names":["VectorStoreRetriever"],"sources":["../../src/retrievers/matryoshka_retriever.ts"],"sourcesContent":["import { DocumentInterface } from \"@langchain/core/documents\";\nimport { Embeddings } from \"@langchain/core/embeddings\";\nimport {\n  cosineSimilarity,\n  euclideanDistance,\n  innerProduct,\n} from \"@langchain/core/utils/math\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\n\n/**\n * Type for options when adding a document to the VectorStore.\n */\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\ntype AddDocumentOptions = Record<string, any>;\n\nexport interface MatryoshkaRetrieverFields {\n  /**\n   * The number of documents to retrieve from the small store.\n   * @default 50\n   */\n  smallK?: number;\n  /**\n   * The number of documents to retrieve from the large store.\n   * @default 8\n   */\n  largeK?: number;\n  /**\n   * The metadata key to store the larger embeddings.\n   * @default \"lc_large_embedding\"\n   */\n  largeEmbeddingKey?: string;\n  /**\n   * The embedding model to use when generating the large\n   * embeddings.\n   */\n  largeEmbeddingModel: Embeddings;\n  /**\n   * The type of search to perform using the large embeddings.\n   * @default \"cosine\"\n   */\n  searchType?: \"cosine\" | \"innerProduct\" | \"euclidean\";\n}\n\n/**\n * A retriever that uses two sets of embeddings to perform adaptive retrieval. Based\n * off of the \"Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval\"\n * blog post {@link https://supabase.com/blog/matryoshka-embeddings}.\n *\n *\n * This class performs \"Adaptive Retrieval\" for searching text embeddings efficiently using the\n * Matryoshka Representation Learning (MRL) technique. It retrieves documents similar to a query\n * embedding in two steps:\n *\n * First-pass: Uses a lower dimensional sub-vector from the MRL embedding for an initial, fast,\n * but less accurate search.\n *\n * Second-pass: Re-ranks the top results from the first pass using the full, high-dimensional\n * embedding for higher accuracy.\n *\n *\n * This code implements MRL embeddings for efficient vector search by combining faster,\n * lower-dimensional initial search with accurate, high-dimensional re-ranking.\n */\nexport class MatryoshkaRetriever<\n  Store extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<Store> {\n  smallK = 50;\n\n  largeK = 8;\n\n  largeEmbeddingKey = \"lc_large_embedding\";\n\n  largeEmbeddingModel: Embeddings;\n\n  searchType: \"cosine\" | \"innerProduct\" | \"euclidean\" = \"cosine\";\n\n  constructor(\n    fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>\n  ) {\n    super(fields);\n    this.smallK = fields.smallK ?? this.smallK;\n    this.largeK = fields.largeK ?? this.largeK;\n    this.largeEmbeddingKey = fields.largeEmbeddingKey ?? this.largeEmbeddingKey;\n    this.largeEmbeddingModel = fields.largeEmbeddingModel;\n    this.searchType = fields.searchType ?? this.searchType;\n  }\n\n  /**\n   * Ranks documents based on their similarity to a query embedding using larger embeddings.\n   *\n   * This method takes a query embedding and a list of documents (smallResults) as input. Each document\n   * in the smallResults array has previously been associated with a large embedding stored in its metadata.\n   * Depending on the `searchType` (cosine, innerProduct, or euclidean), it calculates the similarity scores\n   * between the query embedding and each document's large embedding. It then ranks the documents based on\n   * these similarity scores, from the most similar to the least similar.\n   *\n   * The method returns a promise that resolves to an array of the top `largeK` documents, where `largeK`\n   * is a class property defining the number of documents to return. This subset of documents is determined\n   * by sorting the entire list of documents based on their similarity scores and then selecting the top\n   * `largeK` documents.\n   *\n   * @param {number[]} embeddedQuery The embedding of the query, represented as an array of numbers.\n   * @param {DocumentInterface[]} smallResults An array of documents, each with metadata that includes a large embedding for similarity comparison.\n   * @returns {Promise<DocumentInterface[]>} A promise that resolves to an array of the top `largeK` ranked documents based on their similarity to the query embedding.\n   */\n  private _rankByLargeEmbeddings(\n    embeddedQuery: number[],\n    smallResults: DocumentInterface[]\n  ): DocumentInterface[] {\n    const largeEmbeddings: Array<number[]> = smallResults.map((doc) =>\n      JSON.parse(doc.metadata[this.largeEmbeddingKey])\n    );\n    let func: () => Array<number[]>;\n\n    switch (this.searchType) {\n      case \"cosine\":\n        func = () => cosineSimilarity([embeddedQuery], largeEmbeddings);\n        break;\n      case \"innerProduct\":\n        func = () => innerProduct([embeddedQuery], largeEmbeddings);\n        break;\n      case \"euclidean\":\n        func = () => euclideanDistance([embeddedQuery], largeEmbeddings);\n        break;\n      default:\n        throw new Error(`Unknown search type: ${this.searchType}`);\n    }\n\n    // Calculate the similarity scores between the query embedding and the large embeddings\n    const [similarityScores] = func();\n\n    // Create an array of indices from 0 to N-1, where N is the number of documents\n    let indices = Array.from(\n      { length: smallResults.length },\n      (_, index) => index\n    );\n\n    indices = indices\n      .map((v, i) => [similarityScores[i], v])\n      .sort(([a], [b]) => b - a)\n      .slice(0, this.largeK)\n      .map(([, i]) => i);\n\n    return indices.map((i) => smallResults[i]);\n  }\n\n  async _getRelevantDocuments(query: string): Promise<DocumentInterface[]> {\n    const [embeddedQuery, smallResults] = await Promise.all([\n      this.largeEmbeddingModel.embedQuery(query),\n      this.vectorStore.similaritySearch(query, this.smallK, this.filter),\n    ]);\n\n    return this._rankByLargeEmbeddings(embeddedQuery, smallResults);\n  }\n\n  /**\n   * Override the default `addDocuments` method to embed the documents twice,\n   * once using the larger embeddings model, and then again using the default\n   * embedding model linked to the vector store.\n   *\n   * @param {DocumentInterface[]} documents - An array of documents to add to the vector store.\n   * @param {AddDocumentOptions} options - An optional object containing additional options for adding documents.\n   * @returns {Promise<string[] | void>} A promise that resolves to an array of the document IDs that were added to the vector store.\n   */\n  override addDocuments = async (\n    documents: DocumentInterface[],\n    options?: AddDocumentOptions\n  ): Promise<string[] | void> => {\n    // Insure documents metadata does not contain the large embedding key\n    if (documents.some((doc) => this.largeEmbeddingKey in doc.metadata)) {\n      throw new Error(\n        `All documents must not contain the large embedding key: ${this.largeEmbeddingKey} in their metadata.`\n      );\n    }\n\n    const allDocPageContent = documents.map((doc) => doc.pageContent);\n    const allDocLargeEmbeddings =\n      await this.largeEmbeddingModel.embedDocuments(allDocPageContent);\n\n    const newDocuments: Array<DocumentInterface> = documents.map(\n      (doc, idx) => ({\n        ...doc,\n        metadata: {\n          ...doc.metadata,\n          [this.largeEmbeddingKey]: JSON.stringify(allDocLargeEmbeddings[idx]),\n        },\n      })\n    );\n\n    return this.vectorStore.addDocuments(newDocuments, options);\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAmEA,IAAa,sBAAb,cAEUA,kDAA4B;CACpC,SAAS;CAET,SAAS;CAET,oBAAoB;CAEpB;CAEA,aAAsD;CAEtD,YACE,QACA;AACA,QAAM,OAAO;AACb,OAAK,SAAS,OAAO,UAAU,KAAK;AACpC,OAAK,SAAS,OAAO,UAAU,KAAK;AACpC,OAAK,oBAAoB,OAAO,qBAAqB,KAAK;AAC1D,OAAK,sBAAsB,OAAO;AAClC,OAAK,aAAa,OAAO,cAAc,KAAK;;;;;;;;;;;;;;;;;;;;CAqB9C,AAAQ,uBACN,eACA,cACqB;EACrB,MAAM,kBAAmC,aAAa,KAAK,QACzD,KAAK,MAAM,IAAI,SAAS,KAAK,mBAAmB,CACjD;EACD,IAAI;AAEJ,UAAQ,KAAK,YAAb;GACE,KAAK;AACH,kEAA8B,CAAC,cAAc,EAAE,gBAAgB;AAC/D;GACF,KAAK;AACH,8DAA0B,CAAC,cAAc,EAAE,gBAAgB;AAC3D;GACF,KAAK;AACH,mEAA+B,CAAC,cAAc,EAAE,gBAAgB;AAChE;GACF,QACE,OAAM,IAAI,MAAM,wBAAwB,KAAK,aAAa;;EAI9D,MAAM,CAAC,oBAAoB,MAAM;EAGjC,IAAI,UAAU,MAAM,KAClB,EAAE,QAAQ,aAAa,QAAQ,GAC9B,GAAG,UAAU,MACf;AAED,YAAU,QACP,KAAK,GAAG,MAAM,CAAC,iBAAiB,IAAI,EAAE,CAAC,CACvC,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,EAAE,CACzB,MAAM,GAAG,KAAK,OAAO,CACrB,KAAK,GAAG,OAAO,EAAE;AAEpB,SAAO,QAAQ,KAAK,MAAM,aAAa,GAAG;;CAG5C,MAAM,sBAAsB,OAA6C;EACvE,MAAM,CAAC,eAAe,gBAAgB,MAAM,QAAQ,IAAI,CACtD,KAAK,oBAAoB,WAAW,MAAM,EAC1C,KAAK,YAAY,iBAAiB,OAAO,KAAK,QAAQ,KAAK,OAAO,CACnE,CAAC;AAEF,SAAO,KAAK,uBAAuB,eAAe,aAAa;;;;;;;;;;;CAYjE,AAAS,eAAe,OACtB,WACA,YAC6B;AAE7B,MAAI,UAAU,MAAM,QAAQ,KAAK,qBAAqB,IAAI,SAAS,CACjE,OAAM,IAAI,MACR,2DAA2D,KAAK,kBAAkB,qBACnF;EAGH,MAAM,oBAAoB,UAAU,KAAK,QAAQ,IAAI,YAAY;EACjE,MAAM,wBACJ,MAAM,KAAK,oBAAoB,eAAe,kBAAkB;EAElE,MAAM,eAAyC,UAAU,KACtD,KAAK,SAAS;GACb,GAAG;GACH,UAAU;IACR,GAAG,IAAI;KACN,KAAK,oBAAoB,KAAK,UAAU,sBAAsB,KAAK;IACrE;GACF,EACF;AAED,SAAO,KAAK,YAAY,aAAa,cAAc,QAAQ"}

package/dist/retrievers/matryoshka_retriever.d.cts

@@ -3,7 +3,6 @@ import { DocumentInterface } from "@langchain/core/documents";
 import { Embeddings } from "@langchain/core/embeddings";
 
 //#region src/retrievers/matryoshka_retriever.d.ts
-
 /**
  * Type for options when adding a document to the VectorStore.
  */

package/dist/retrievers/matryoshka_retriever.d.cts.map

@@ -1 +1 @@
-{"version":3,"file":"matryoshka_retriever.d.cts","names":["DocumentInterface","Embeddings","VectorStore","VectorStoreRetriever","VectorStoreRetrieverInput","AddDocumentOptions","Record","MatryoshkaRetrieverFields","MatryoshkaRetriever","Store","Promise"],"sources":["../../src/retrievers/matryoshka_retriever.d.ts"],"sourcesContent":["import { DocumentInterface } from \"@langchain/core/documents\";\nimport { Embeddings } from \"@langchain/core/embeddings\";\nimport { VectorStore, VectorStoreRetriever, VectorStoreRetrieverInput } from \"@langchain/core/vectorstores\";\n/**\n * Type for options when adding a document to the VectorStore.\n */\ntype AddDocumentOptions = Record<string, any>;\nexport interface MatryoshkaRetrieverFields {\n    /**\n     * The number of documents to retrieve from the small store.\n     * @default 50\n     */\n    smallK?: number;\n    /**\n     * The number of documents to retrieve from the large store.\n     * @default 8\n     */\n    largeK?: number;\n    /**\n     * The metadata key to store the larger embeddings.\n     * @default \"lc_large_embedding\"\n     */\n    largeEmbeddingKey?: string;\n    /**\n     * The embedding model to use when generating the large\n     * embeddings.\n     */\n    largeEmbeddingModel: Embeddings;\n    /**\n     * The type of search to perform using the large embeddings.\n     * @default \"cosine\"\n     */\n    searchType?: \"cosine\" | \"innerProduct\" | \"euclidean\";\n}\n/**\n * A retriever that uses two sets of embeddings to perform adaptive retrieval. Based\n * off of the \"Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval\"\n * blog post {@link https://supabase.com/blog/matryoshka-embeddings}.\n *\n *\n * This class performs \"Adaptive Retrieval\" for searching text embeddings efficiently using the\n * Matryoshka Representation Learning (MRL) technique. It retrieves documents similar to a query\n * embedding in two steps:\n *\n * First-pass: Uses a lower dimensional sub-vector from the MRL embedding for an initial, fast,\n * but less accurate search.\n *\n * Second-pass: Re-ranks the top results from the first pass using the full, high-dimensional\n * embedding for higher accuracy.\n *\n *\n * This code implements MRL embeddings for efficient vector search by combining faster,\n * lower-dimensional initial search with accurate, high-dimensional re-ranking.\n */\nexport declare class MatryoshkaRetriever<Store extends VectorStore = VectorStore> extends VectorStoreRetriever<Store> {\n    smallK: number;\n    largeK: number;\n    largeEmbeddingKey: string;\n    largeEmbeddingModel: Embeddings;\n    searchType: \"cosine\" | \"innerProduct\" | \"euclidean\";\n    constructor(fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>);\n    /**\n     * Ranks documents based on their similarity to a query embedding using larger embeddings.\n     *\n     * This method takes a query embedding and a list of documents (smallResults) as input. Each document\n     * in the smallResults array has previously been associated with a large embedding stored in its metadata.\n     * Depending on the `searchType` (cosine, innerProduct, or euclidean), it calculates the similarity scores\n     * between the query embedding and each document's large embedding. It then ranks the documents based on\n     * these similarity scores, from the most similar to the least similar.\n     *\n     * The method returns a promise that resolves to an array of the top `largeK` documents, where `largeK`\n     * is a class property defining the number of documents to return. This subset of documents is determined\n     * by sorting the entire list of documents based on their similarity scores and then selecting the top\n     * `largeK` documents.\n     *\n     * @param {number[]} embeddedQuery The embedding of the query, represented as an array of numbers.\n     * @param {DocumentInterface[]} smallResults An array of documents, each with metadata that includes a large embedding for similarity comparison.\n     * @returns {Promise<DocumentInterface[]>} A promise that resolves to an array of the top `largeK` ranked documents based on their similarity to the query embedding.\n     */\n    private _rankByLargeEmbeddings;\n    _getRelevantDocuments(query: string): Promise<DocumentInterface[]>;\n    /**\n     * Override the default `addDocuments` method to embed the documents twice,\n     * once using the larger embeddings model, and then again using the default\n     * embedding model linked to the vector store.\n     *\n     * @param {DocumentInterface[]} documents - An array of documents to add to the vector store.\n     * @param {AddDocumentOptions} options - An optional object containing additional options for adding documents.\n     * @returns {Promise<string[] | void>} A promise that resolves to an array of the document IDs that were added to the vector store.\n     */\n    addDocuments: (documents: DocumentInterface<Record<string, any>>[], options?: AddDocumentOptions | undefined) => Promise<void | string[]>;\n}\nexport {};\n//# sourceMappingURL=matryoshka_retriever.d.ts.map"],"mappings":";;;;;;;;AAE4G;AAK5G,KADKK,kBAAAA,GAAqBC,MACgB,CAAA,MAAA,EAAA,GAoBjBL,CAAAA;AA2BJO,UA/CJD,yBAAAA,CA+CuB;EAAeL;;;;EAM/BK,MAAAA,CAAAA,EAAAA,MAAAA;EAAsDE;;;;EA8B9BH,MAAAA,CAAAA,EAAAA,MAAAA;EAAlBN;;;;EApCgF,iBAAA,CAAA,EAAA,MAAA;;;;;uBA3BrFC;;;;;;;;;;;;;;;;;;;;;;;;;;;cA2BJO,kCAAkCN,cAAcA,qBAAqBC,qBAAqBM;;;;uBAItFR;;sBAEDM,4BAA4BH,0BAA0BK;;;;;;;;;;;;;;;;;;;;wCAoBpCC,QAAQV;;;;;;;;;;4BAUpBA,kBAAkBM,kCAAkCD,mCAAmCK"}
+{"version":3,"file":"matryoshka_retriever.d.cts","names":[],"sources":["../../src/retrievers/matryoshka_retriever.ts"],"mappings":";;;;;;;AAWsC;KAMjC,kBAAA,GAAqB,MAAA;AAAA,UAET,yBAAA;EAFS;;AAE1B;;EAKE,MAAA;EAe+B;;;;EAV/B,MAAA;EAUqB;;;;EALrB,iBAAA;EAiC8B;;;;EA5B9B,mBAAA,EAAqB,UAAA;EAqCA;;;;EAhCrB,UAAA;AAAA;;;;;;;;;;;;;;;;;;;;;cAuBW,mBAAA,eACG,WAAA,GAAc,WAAA,UACpB,oBAAA,CAAqB,KAAA;EAC7B,MAAA;EAEA,MAAA;EAEA,iBAAA;EAEA,mBAAA,EAAqB,UAAA;EAErB,UAAA;EAEA,WAAA,CACE,MAAA,EAAQ,yBAAA,GAA4B,yBAAA,CAA0B,KAAA;EAqEZ;;;;;;;;;;;;;;;;;;EAAA,QAzC5C,sBAAA;EAyCF,qBAAA,CAAsB,KAAA,WAAgB,OAAA,CAAQ,iBAAA;;;;;;;;;;EAkB3C,YAAA,GAAY,SAAA,EAAA,iBAAA,CAAA,MAAA,kBAAA,OAAA,GAAA,kBAAA,iBAAA,OAAA;AAAA"}

package/dist/retrievers/matryoshka_retriever.d.ts

@@ -3,7 +3,6 @@ import { Embeddings } from "@langchain/core/embeddings";
 import { VectorStore, VectorStoreRetriever, VectorStoreRetrieverInput } from "@langchain/core/vectorstores";
 
 //#region src/retrievers/matryoshka_retriever.d.ts
-
 /**
  * Type for options when adding a document to the VectorStore.
  */

package/dist/retrievers/matryoshka_retriever.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"matryoshka_retriever.d.ts","names":["DocumentInterface","Embeddings","VectorStore","VectorStoreRetriever","VectorStoreRetrieverInput","AddDocumentOptions","Record","MatryoshkaRetrieverFields","MatryoshkaRetriever","Store","Promise"],"sources":["../../src/retrievers/matryoshka_retriever.d.ts"],"sourcesContent":["import { DocumentInterface } from \"@langchain/core/documents\";\nimport { Embeddings } from \"@langchain/core/embeddings\";\nimport { VectorStore, VectorStoreRetriever, VectorStoreRetrieverInput } from \"@langchain/core/vectorstores\";\n/**\n * Type for options when adding a document to the VectorStore.\n */\ntype AddDocumentOptions = Record<string, any>;\nexport interface MatryoshkaRetrieverFields {\n    /**\n     * The number of documents to retrieve from the small store.\n     * @default 50\n     */\n    smallK?: number;\n    /**\n     * The number of documents to retrieve from the large store.\n     * @default 8\n     */\n    largeK?: number;\n    /**\n     * The metadata key to store the larger embeddings.\n     * @default \"lc_large_embedding\"\n     */\n    largeEmbeddingKey?: string;\n    /**\n     * The embedding model to use when generating the large\n     * embeddings.\n     */\n    largeEmbeddingModel: Embeddings;\n    /**\n     * The type of search to perform using the large embeddings.\n     * @default \"cosine\"\n     */\n    searchType?: \"cosine\" | \"innerProduct\" | \"euclidean\";\n}\n/**\n * A retriever that uses two sets of embeddings to perform adaptive retrieval. Based\n * off of the \"Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval\"\n * blog post {@link https://supabase.com/blog/matryoshka-embeddings}.\n *\n *\n * This class performs \"Adaptive Retrieval\" for searching text embeddings efficiently using the\n * Matryoshka Representation Learning (MRL) technique. It retrieves documents similar to a query\n * embedding in two steps:\n *\n * First-pass: Uses a lower dimensional sub-vector from the MRL embedding for an initial, fast,\n * but less accurate search.\n *\n * Second-pass: Re-ranks the top results from the first pass using the full, high-dimensional\n * embedding for higher accuracy.\n *\n *\n * This code implements MRL embeddings for efficient vector search by combining faster,\n * lower-dimensional initial search with accurate, high-dimensional re-ranking.\n */\nexport declare class MatryoshkaRetriever<Store extends VectorStore = VectorStore> extends VectorStoreRetriever<Store> {\n    smallK: number;\n    largeK: number;\n    largeEmbeddingKey: string;\n    largeEmbeddingModel: Embeddings;\n    searchType: \"cosine\" | \"innerProduct\" | \"euclidean\";\n    constructor(fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>);\n    /**\n     * Ranks documents based on their similarity to a query embedding using larger embeddings.\n     *\n     * This method takes a query embedding and a list of documents (smallResults) as input. Each document\n     * in the smallResults array has previously been associated with a large embedding stored in its metadata.\n     * Depending on the `searchType` (cosine, innerProduct, or euclidean), it calculates the similarity scores\n     * between the query embedding and each document's large embedding. It then ranks the documents based on\n     * these similarity scores, from the most similar to the least similar.\n     *\n     * The method returns a promise that resolves to an array of the top `largeK` documents, where `largeK`\n     * is a class property defining the number of documents to return. This subset of documents is determined\n     * by sorting the entire list of documents based on their similarity scores and then selecting the top\n     * `largeK` documents.\n     *\n     * @param {number[]} embeddedQuery The embedding of the query, represented as an array of numbers.\n     * @param {DocumentInterface[]} smallResults An array of documents, each with metadata that includes a large embedding for similarity comparison.\n     * @returns {Promise<DocumentInterface[]>} A promise that resolves to an array of the top `largeK` ranked documents based on their similarity to the query embedding.\n     */\n    private _rankByLargeEmbeddings;\n    _getRelevantDocuments(query: string): Promise<DocumentInterface[]>;\n    /**\n     * Override the default `addDocuments` method to embed the documents twice,\n     * once using the larger embeddings model, and then again using the default\n     * embedding model linked to the vector store.\n     *\n     * @param {DocumentInterface[]} documents - An array of documents to add to the vector store.\n     * @param {AddDocumentOptions} options - An optional object containing additional options for adding documents.\n     * @returns {Promise<string[] | void>} A promise that resolves to an array of the document IDs that were added to the vector store.\n     */\n    addDocuments: (documents: DocumentInterface<Record<string, any>>[], options?: AddDocumentOptions | undefined) => Promise<void | string[]>;\n}\nexport {};\n//# sourceMappingURL=matryoshka_retriever.d.ts.map"],"mappings":";;;;;;;;AAE4G;AAK5G,KADKK,kBAAAA,GAAqBC,MACgB,CAAA,MAAA,EAAA,GAoBjBL,CAAAA;AA2BJO,UA/CJD,yBAAAA,CA+CuB;EAAeL;;;;EAM/BK,MAAAA,CAAAA,EAAAA,MAAAA;EAAsDE;;;;EA8B9BH,MAAAA,CAAAA,EAAAA,MAAAA;EAAlBN;;;;EApCgF,iBAAA,CAAA,EAAA,MAAA;;;;;uBA3BrFC;;;;;;;;;;;;;;;;;;;;;;;;;;;cA2BJO,kCAAkCN,cAAcA,qBAAqBC,qBAAqBM;;;;uBAItFR;;sBAEDM,4BAA4BH,0BAA0BK;;;;;;;;;;;;;;;;;;;;wCAoBpCC,QAAQV;;;;;;;;;;4BAUpBA,kBAAkBM,kCAAkCD,mCAAmCK"}
+{"version":3,"file":"matryoshka_retriever.d.ts","names":[],"sources":["../../src/retrievers/matryoshka_retriever.ts"],"mappings":";;;;;;;AAWsC;KAMjC,kBAAA,GAAqB,MAAA;AAAA,UAET,yBAAA;EAFS;;AAE1B;;EAKE,MAAA;EAe+B;;;;EAV/B,MAAA;EAUqB;;;;EALrB,iBAAA;EAiC8B;;;;EA5B9B,mBAAA,EAAqB,UAAA;EAqCA;;;;EAhCrB,UAAA;AAAA;;;;;;;;;;;;;;;;;;;;;cAuBW,mBAAA,eACG,WAAA,GAAc,WAAA,UACpB,oBAAA,CAAqB,KAAA;EAC7B,MAAA;EAEA,MAAA;EAEA,iBAAA;EAEA,mBAAA,EAAqB,UAAA;EAErB,UAAA;EAEA,WAAA,CACE,MAAA,EAAQ,yBAAA,GAA4B,yBAAA,CAA0B,KAAA;EAqEZ;;;;;;;;;;;;;;;;;;EAAA,QAzC5C,sBAAA;EAyCF,qBAAA,CAAsB,KAAA,WAAgB,OAAA,CAAQ,iBAAA;;;;;;;;;;EAkB3C,YAAA,GAAY,SAAA,EAAA,iBAAA,CAAA,MAAA,kBAAA,OAAA,GAAA,kBAAA,iBAAA,OAAA;AAAA"}

package/dist/retrievers/matryoshka_retriever.js

@@ -1,10 +1,9 @@
-import { __export } from "../_virtual/rolldown_runtime.js";
+import { __exportAll } from "../_virtual/_rolldown/runtime.js";
 import { cosineSimilarity, euclideanDistance, innerProduct } from "@langchain/core/utils/math";
 import { VectorStoreRetriever } from "@langchain/core/vectorstores";
 
 //#region src/retrievers/matryoshka_retriever.ts
-var matryoshka_retriever_exports = {};
-__export(matryoshka_retriever_exports, { MatryoshkaRetriever: () => MatryoshkaRetriever });
+var matryoshka_retriever_exports = /* @__PURE__ */ __exportAll({ MatryoshkaRetriever: () => MatryoshkaRetriever });
 /**
 * A retriever that uses two sets of embeddings to perform adaptive retrieval. Based
 * off of the "Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval"

package/dist/retrievers/matryoshka_retriever.js.map

@@ -1 +1 @@
-{"version":3,"file":"matryoshka_retriever.js","names":["fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>","embeddedQuery: number[]","smallResults: DocumentInterface[]","largeEmbeddings: Array<number[]>","func: () => Array<number[]>","query: string","documents: DocumentInterface[]","options?: AddDocumentOptions","newDocuments: Array<DocumentInterface>"],"sources":["../../src/retrievers/matryoshka_retriever.ts"],"sourcesContent":["import { DocumentInterface } from \"@langchain/core/documents\";\nimport { Embeddings } from \"@langchain/core/embeddings\";\nimport {\n  cosineSimilarity,\n  euclideanDistance,\n  innerProduct,\n} from \"@langchain/core/utils/math\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\n\n/**\n * Type for options when adding a document to the VectorStore.\n */\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\ntype AddDocumentOptions = Record<string, any>;\n\nexport interface MatryoshkaRetrieverFields {\n  /**\n   * The number of documents to retrieve from the small store.\n   * @default 50\n   */\n  smallK?: number;\n  /**\n   * The number of documents to retrieve from the large store.\n   * @default 8\n   */\n  largeK?: number;\n  /**\n   * The metadata key to store the larger embeddings.\n   * @default \"lc_large_embedding\"\n   */\n  largeEmbeddingKey?: string;\n  /**\n   * The embedding model to use when generating the large\n   * embeddings.\n   */\n  largeEmbeddingModel: Embeddings;\n  /**\n   * The type of search to perform using the large embeddings.\n   * @default \"cosine\"\n   */\n  searchType?: \"cosine\" | \"innerProduct\" | \"euclidean\";\n}\n\n/**\n * A retriever that uses two sets of embeddings to perform adaptive retrieval. Based\n * off of the \"Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval\"\n * blog post {@link https://supabase.com/blog/matryoshka-embeddings}.\n *\n *\n * This class performs \"Adaptive Retrieval\" for searching text embeddings efficiently using the\n * Matryoshka Representation Learning (MRL) technique. It retrieves documents similar to a query\n * embedding in two steps:\n *\n * First-pass: Uses a lower dimensional sub-vector from the MRL embedding for an initial, fast,\n * but less accurate search.\n *\n * Second-pass: Re-ranks the top results from the first pass using the full, high-dimensional\n * embedding for higher accuracy.\n *\n *\n * This code implements MRL embeddings for efficient vector search by combining faster,\n * lower-dimensional initial search with accurate, high-dimensional re-ranking.\n */\nexport class MatryoshkaRetriever<\n  Store extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<Store> {\n  smallK = 50;\n\n  largeK = 8;\n\n  largeEmbeddingKey = \"lc_large_embedding\";\n\n  largeEmbeddingModel: Embeddings;\n\n  searchType: \"cosine\" | \"innerProduct\" | \"euclidean\" = \"cosine\";\n\n  constructor(\n    fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>\n  ) {\n    super(fields);\n    this.smallK = fields.smallK ?? this.smallK;\n    this.largeK = fields.largeK ?? this.largeK;\n    this.largeEmbeddingKey = fields.largeEmbeddingKey ?? this.largeEmbeddingKey;\n    this.largeEmbeddingModel = fields.largeEmbeddingModel;\n    this.searchType = fields.searchType ?? this.searchType;\n  }\n\n  /**\n   * Ranks documents based on their similarity to a query embedding using larger embeddings.\n   *\n   * This method takes a query embedding and a list of documents (smallResults) as input. Each document\n   * in the smallResults array has previously been associated with a large embedding stored in its metadata.\n   * Depending on the `searchType` (cosine, innerProduct, or euclidean), it calculates the similarity scores\n   * between the query embedding and each document's large embedding. It then ranks the documents based on\n   * these similarity scores, from the most similar to the least similar.\n   *\n   * The method returns a promise that resolves to an array of the top `largeK` documents, where `largeK`\n   * is a class property defining the number of documents to return. This subset of documents is determined\n   * by sorting the entire list of documents based on their similarity scores and then selecting the top\n   * `largeK` documents.\n   *\n   * @param {number[]} embeddedQuery The embedding of the query, represented as an array of numbers.\n   * @param {DocumentInterface[]} smallResults An array of documents, each with metadata that includes a large embedding for similarity comparison.\n   * @returns {Promise<DocumentInterface[]>} A promise that resolves to an array of the top `largeK` ranked documents based on their similarity to the query embedding.\n   */\n  private _rankByLargeEmbeddings(\n    embeddedQuery: number[],\n    smallResults: DocumentInterface[]\n  ): DocumentInterface[] {\n    const largeEmbeddings: Array<number[]> = smallResults.map((doc) =>\n      JSON.parse(doc.metadata[this.largeEmbeddingKey])\n    );\n    let func: () => Array<number[]>;\n\n    switch (this.searchType) {\n      case \"cosine\":\n        func = () => cosineSimilarity([embeddedQuery], largeEmbeddings);\n        break;\n      case \"innerProduct\":\n        func = () => innerProduct([embeddedQuery], largeEmbeddings);\n        break;\n      case \"euclidean\":\n        func = () => euclideanDistance([embeddedQuery], largeEmbeddings);\n        break;\n      default:\n        throw new Error(`Unknown search type: ${this.searchType}`);\n    }\n\n    // Calculate the similarity scores between the query embedding and the large embeddings\n    const [similarityScores] = func();\n\n    // Create an array of indices from 0 to N-1, where N is the number of documents\n    let indices = Array.from(\n      { length: smallResults.length },\n      (_, index) => index\n    );\n\n    indices = indices\n      .map((v, i) => [similarityScores[i], v])\n      .sort(([a], [b]) => b - a)\n      .slice(0, this.largeK)\n      .map(([, i]) => i);\n\n    return indices.map((i) => smallResults[i]);\n  }\n\n  async _getRelevantDocuments(query: string): Promise<DocumentInterface[]> {\n    const [embeddedQuery, smallResults] = await Promise.all([\n      this.largeEmbeddingModel.embedQuery(query),\n      this.vectorStore.similaritySearch(query, this.smallK, this.filter),\n    ]);\n\n    return this._rankByLargeEmbeddings(embeddedQuery, smallResults);\n  }\n\n  /**\n   * Override the default `addDocuments` method to embed the documents twice,\n   * once using the larger embeddings model, and then again using the default\n   * embedding model linked to the vector store.\n   *\n   * @param {DocumentInterface[]} documents - An array of documents to add to the vector store.\n   * @param {AddDocumentOptions} options - An optional object containing additional options for adding documents.\n   * @returns {Promise<string[] | void>} A promise that resolves to an array of the document IDs that were added to the vector store.\n   */\n  override addDocuments = async (\n    documents: DocumentInterface[],\n    options?: AddDocumentOptions\n  ): Promise<string[] | void> => {\n    // Insure documents metadata does not contain the large embedding key\n    if (documents.some((doc) => this.largeEmbeddingKey in doc.metadata)) {\n      throw new Error(\n        `All documents must not contain the large embedding key: ${this.largeEmbeddingKey} in their metadata.`\n      );\n    }\n\n    const allDocPageContent = documents.map((doc) => doc.pageContent);\n    const allDocLargeEmbeddings =\n      await this.largeEmbeddingModel.embedDocuments(allDocPageContent);\n\n    const newDocuments: Array<DocumentInterface> = documents.map(\n      (doc, idx) => ({\n        ...doc,\n        metadata: {\n          ...doc.metadata,\n          [this.largeEmbeddingKey]: JSON.stringify(allDocLargeEmbeddings[idx]),\n        },\n      })\n    );\n\n    return this.vectorStore.addDocuments(newDocuments, options);\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAmEA,IAAa,sBAAb,cAEU,qBAA4B;CACpC,SAAS;CAET,SAAS;CAET,oBAAoB;CAEpB;CAEA,aAAsD;CAEtD,YACEA,QACA;EACA,MAAM,OAAO;EACb,KAAK,SAAS,OAAO,UAAU,KAAK;EACpC,KAAK,SAAS,OAAO,UAAU,KAAK;EACpC,KAAK,oBAAoB,OAAO,qBAAqB,KAAK;EAC1D,KAAK,sBAAsB,OAAO;EAClC,KAAK,aAAa,OAAO,cAAc,KAAK;CAC7C;;;;;;;;;;;;;;;;;;;CAoBD,AAAQ,uBACNC,eACAC,cACqB;EACrB,MAAMC,kBAAmC,aAAa,IAAI,CAAC,QACzD,KAAK,MAAM,IAAI,SAAS,KAAK,mBAAmB,CACjD;EACD,IAAIC;AAEJ,UAAQ,KAAK,YAAb;GACE,KAAK;IACH,OAAO,MAAM,iBAAiB,CAAC,aAAc,GAAE,gBAAgB;AAC/D;GACF,KAAK;IACH,OAAO,MAAM,aAAa,CAAC,aAAc,GAAE,gBAAgB;AAC3D;GACF,KAAK;IACH,OAAO,MAAM,kBAAkB,CAAC,aAAc,GAAE,gBAAgB;AAChE;GACF,QACE,OAAM,IAAI,MAAM,CAAC,qBAAqB,EAAE,KAAK,YAAY;EAC5D;EAGD,MAAM,CAAC,iBAAiB,GAAG,MAAM;EAGjC,IAAI,UAAU,MAAM,KAClB,EAAE,QAAQ,aAAa,OAAQ,GAC/B,CAAC,GAAG,UAAU,MACf;EAED,UAAU,QACP,IAAI,CAAC,GAAG,MAAM,CAAC,iBAAiB,IAAI,CAAE,EAAC,CACvC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,KAAK,IAAI,EAAE,CACzB,MAAM,GAAG,KAAK,OAAO,CACrB,IAAI,CAAC,GAAG,EAAE,KAAK,EAAE;AAEpB,SAAO,QAAQ,IAAI,CAAC,MAAM,aAAa,GAAG;CAC3C;CAED,MAAM,sBAAsBC,OAA6C;EACvE,MAAM,CAAC,eAAe,aAAa,GAAG,MAAM,QAAQ,IAAI,CACtD,KAAK,oBAAoB,WAAW,MAAM,EAC1C,KAAK,YAAY,iBAAiB,OAAO,KAAK,QAAQ,KAAK,OAAO,AACnE,EAAC;AAEF,SAAO,KAAK,uBAAuB,eAAe,aAAa;CAChE;;;;;;;;;;CAWD,AAAS,eAAe,OACtBC,WACAC,YAC6B;AAE7B,MAAI,UAAU,KAAK,CAAC,QAAQ,KAAK,qBAAqB,IAAI,SAAS,CACjE,OAAM,IAAI,MACR,CAAC,wDAAwD,EAAE,KAAK,kBAAkB,mBAAmB,CAAC;EAI1G,MAAM,oBAAoB,UAAU,IAAI,CAAC,QAAQ,IAAI,YAAY;EACjE,MAAM,wBACJ,MAAM,KAAK,oBAAoB,eAAe,kBAAkB;EAElE,MAAMC,eAAyC,UAAU,IACvD,CAAC,KAAK,SAAS;GACb,GAAG;GACH,UAAU;IACR,GAAG,IAAI;KACN,KAAK,oBAAoB,KAAK,UAAU,sBAAsB,KAAK;GACrE;EACF,GACF;AAED,SAAO,KAAK,YAAY,aAAa,cAAc,QAAQ;CAC5D;AACF"}
+{"version":3,"file":"matryoshka_retriever.js","names":[],"sources":["../../src/retrievers/matryoshka_retriever.ts"],"sourcesContent":["import { DocumentInterface } from \"@langchain/core/documents\";\nimport { Embeddings } from \"@langchain/core/embeddings\";\nimport {\n  cosineSimilarity,\n  euclideanDistance,\n  innerProduct,\n} from \"@langchain/core/utils/math\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\n\n/**\n * Type for options when adding a document to the VectorStore.\n */\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\ntype AddDocumentOptions = Record<string, any>;\n\nexport interface MatryoshkaRetrieverFields {\n  /**\n   * The number of documents to retrieve from the small store.\n   * @default 50\n   */\n  smallK?: number;\n  /**\n   * The number of documents to retrieve from the large store.\n   * @default 8\n   */\n  largeK?: number;\n  /**\n   * The metadata key to store the larger embeddings.\n   * @default \"lc_large_embedding\"\n   */\n  largeEmbeddingKey?: string;\n  /**\n   * The embedding model to use when generating the large\n   * embeddings.\n   */\n  largeEmbeddingModel: Embeddings;\n  /**\n   * The type of search to perform using the large embeddings.\n   * @default \"cosine\"\n   */\n  searchType?: \"cosine\" | \"innerProduct\" | \"euclidean\";\n}\n\n/**\n * A retriever that uses two sets of embeddings to perform adaptive retrieval. Based\n * off of the \"Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval\"\n * blog post {@link https://supabase.com/blog/matryoshka-embeddings}.\n *\n *\n * This class performs \"Adaptive Retrieval\" for searching text embeddings efficiently using the\n * Matryoshka Representation Learning (MRL) technique. It retrieves documents similar to a query\n * embedding in two steps:\n *\n * First-pass: Uses a lower dimensional sub-vector from the MRL embedding for an initial, fast,\n * but less accurate search.\n *\n * Second-pass: Re-ranks the top results from the first pass using the full, high-dimensional\n * embedding for higher accuracy.\n *\n *\n * This code implements MRL embeddings for efficient vector search by combining faster,\n * lower-dimensional initial search with accurate, high-dimensional re-ranking.\n */\nexport class MatryoshkaRetriever<\n  Store extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<Store> {\n  smallK = 50;\n\n  largeK = 8;\n\n  largeEmbeddingKey = \"lc_large_embedding\";\n\n  largeEmbeddingModel: Embeddings;\n\n  searchType: \"cosine\" | \"innerProduct\" | \"euclidean\" = \"cosine\";\n\n  constructor(\n    fields: MatryoshkaRetrieverFields & VectorStoreRetrieverInput<Store>\n  ) {\n    super(fields);\n    this.smallK = fields.smallK ?? this.smallK;\n    this.largeK = fields.largeK ?? this.largeK;\n    this.largeEmbeddingKey = fields.largeEmbeddingKey ?? this.largeEmbeddingKey;\n    this.largeEmbeddingModel = fields.largeEmbeddingModel;\n    this.searchType = fields.searchType ?? this.searchType;\n  }\n\n  /**\n   * Ranks documents based on their similarity to a query embedding using larger embeddings.\n   *\n   * This method takes a query embedding and a list of documents (smallResults) as input. Each document\n   * in the smallResults array has previously been associated with a large embedding stored in its metadata.\n   * Depending on the `searchType` (cosine, innerProduct, or euclidean), it calculates the similarity scores\n   * between the query embedding and each document's large embedding. It then ranks the documents based on\n   * these similarity scores, from the most similar to the least similar.\n   *\n   * The method returns a promise that resolves to an array of the top `largeK` documents, where `largeK`\n   * is a class property defining the number of documents to return. This subset of documents is determined\n   * by sorting the entire list of documents based on their similarity scores and then selecting the top\n   * `largeK` documents.\n   *\n   * @param {number[]} embeddedQuery The embedding of the query, represented as an array of numbers.\n   * @param {DocumentInterface[]} smallResults An array of documents, each with metadata that includes a large embedding for similarity comparison.\n   * @returns {Promise<DocumentInterface[]>} A promise that resolves to an array of the top `largeK` ranked documents based on their similarity to the query embedding.\n   */\n  private _rankByLargeEmbeddings(\n    embeddedQuery: number[],\n    smallResults: DocumentInterface[]\n  ): DocumentInterface[] {\n    const largeEmbeddings: Array<number[]> = smallResults.map((doc) =>\n      JSON.parse(doc.metadata[this.largeEmbeddingKey])\n    );\n    let func: () => Array<number[]>;\n\n    switch (this.searchType) {\n      case \"cosine\":\n        func = () => cosineSimilarity([embeddedQuery], largeEmbeddings);\n        break;\n      case \"innerProduct\":\n        func = () => innerProduct([embeddedQuery], largeEmbeddings);\n        break;\n      case \"euclidean\":\n        func = () => euclideanDistance([embeddedQuery], largeEmbeddings);\n        break;\n      default:\n        throw new Error(`Unknown search type: ${this.searchType}`);\n    }\n\n    // Calculate the similarity scores between the query embedding and the large embeddings\n    const [similarityScores] = func();\n\n    // Create an array of indices from 0 to N-1, where N is the number of documents\n    let indices = Array.from(\n      { length: smallResults.length },\n      (_, index) => index\n    );\n\n    indices = indices\n      .map((v, i) => [similarityScores[i], v])\n      .sort(([a], [b]) => b - a)\n      .slice(0, this.largeK)\n      .map(([, i]) => i);\n\n    return indices.map((i) => smallResults[i]);\n  }\n\n  async _getRelevantDocuments(query: string): Promise<DocumentInterface[]> {\n    const [embeddedQuery, smallResults] = await Promise.all([\n      this.largeEmbeddingModel.embedQuery(query),\n      this.vectorStore.similaritySearch(query, this.smallK, this.filter),\n    ]);\n\n    return this._rankByLargeEmbeddings(embeddedQuery, smallResults);\n  }\n\n  /**\n   * Override the default `addDocuments` method to embed the documents twice,\n   * once using the larger embeddings model, and then again using the default\n   * embedding model linked to the vector store.\n   *\n   * @param {DocumentInterface[]} documents - An array of documents to add to the vector store.\n   * @param {AddDocumentOptions} options - An optional object containing additional options for adding documents.\n   * @returns {Promise<string[] | void>} A promise that resolves to an array of the document IDs that were added to the vector store.\n   */\n  override addDocuments = async (\n    documents: DocumentInterface[],\n    options?: AddDocumentOptions\n  ): Promise<string[] | void> => {\n    // Insure documents metadata does not contain the large embedding key\n    if (documents.some((doc) => this.largeEmbeddingKey in doc.metadata)) {\n      throw new Error(\n        `All documents must not contain the large embedding key: ${this.largeEmbeddingKey} in their metadata.`\n      );\n    }\n\n    const allDocPageContent = documents.map((doc) => doc.pageContent);\n    const allDocLargeEmbeddings =\n      await this.largeEmbeddingModel.embedDocuments(allDocPageContent);\n\n    const newDocuments: Array<DocumentInterface> = documents.map(\n      (doc, idx) => ({\n        ...doc,\n        metadata: {\n          ...doc.metadata,\n          [this.largeEmbeddingKey]: JSON.stringify(allDocLargeEmbeddings[idx]),\n        },\n      })\n    );\n\n    return this.vectorStore.addDocuments(newDocuments, options);\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AAmEA,IAAa,sBAAb,cAEU,qBAA4B;CACpC,SAAS;CAET,SAAS;CAET,oBAAoB;CAEpB;CAEA,aAAsD;CAEtD,YACE,QACA;AACA,QAAM,OAAO;AACb,OAAK,SAAS,OAAO,UAAU,KAAK;AACpC,OAAK,SAAS,OAAO,UAAU,KAAK;AACpC,OAAK,oBAAoB,OAAO,qBAAqB,KAAK;AAC1D,OAAK,sBAAsB,OAAO;AAClC,OAAK,aAAa,OAAO,cAAc,KAAK;;;;;;;;;;;;;;;;;;;;CAqB9C,AAAQ,uBACN,eACA,cACqB;EACrB,MAAM,kBAAmC,aAAa,KAAK,QACzD,KAAK,MAAM,IAAI,SAAS,KAAK,mBAAmB,CACjD;EACD,IAAI;AAEJ,UAAQ,KAAK,YAAb;GACE,KAAK;AACH,iBAAa,iBAAiB,CAAC,cAAc,EAAE,gBAAgB;AAC/D;GACF,KAAK;AACH,iBAAa,aAAa,CAAC,cAAc,EAAE,gBAAgB;AAC3D;GACF,KAAK;AACH,iBAAa,kBAAkB,CAAC,cAAc,EAAE,gBAAgB;AAChE;GACF,QACE,OAAM,IAAI,MAAM,wBAAwB,KAAK,aAAa;;EAI9D,MAAM,CAAC,oBAAoB,MAAM;EAGjC,IAAI,UAAU,MAAM,KAClB,EAAE,QAAQ,aAAa,QAAQ,GAC9B,GAAG,UAAU,MACf;AAED,YAAU,QACP,KAAK,GAAG,MAAM,CAAC,iBAAiB,IAAI,EAAE,CAAC,CACvC,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,EAAE,CACzB,MAAM,GAAG,KAAK,OAAO,CACrB,KAAK,GAAG,OAAO,EAAE;AAEpB,SAAO,QAAQ,KAAK,MAAM,aAAa,GAAG;;CAG5C,MAAM,sBAAsB,OAA6C;EACvE,MAAM,CAAC,eAAe,gBAAgB,MAAM,QAAQ,IAAI,CACtD,KAAK,oBAAoB,WAAW,MAAM,EAC1C,KAAK,YAAY,iBAAiB,OAAO,KAAK,QAAQ,KAAK,OAAO,CACnE,CAAC;AAEF,SAAO,KAAK,uBAAuB,eAAe,aAAa;;;;;;;;;;;CAYjE,AAAS,eAAe,OACtB,WACA,YAC6B;AAE7B,MAAI,UAAU,MAAM,QAAQ,KAAK,qBAAqB,IAAI,SAAS,CACjE,OAAM,IAAI,MACR,2DAA2D,KAAK,kBAAkB,qBACnF;EAGH,MAAM,oBAAoB,UAAU,KAAK,QAAQ,IAAI,YAAY;EACjE,MAAM,wBACJ,MAAM,KAAK,oBAAoB,eAAe,kBAAkB;EAElE,MAAM,eAAyC,UAAU,KACtD,KAAK,SAAS;GACb,GAAG;GACH,UAAU;IACR,GAAG,IAAI;KACN,KAAK,oBAAoB,KAAK,UAAU,sBAAsB,KAAK;IACrE;GACF,EACF;AAED,SAAO,KAAK,YAAY,aAAa,cAAc,QAAQ"}

package/dist/retrievers/multi_query.cjs

@@ -1,13 +1,13 @@
-const require_rolldown_runtime = require('../_virtual/rolldown_runtime.cjs');
+Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });
+const require_runtime = require('../_virtual/_rolldown/runtime.cjs');
 const require_llm_chain = require('../chains/llm_chain.cjs');
-const __langchain_core_prompts = require_rolldown_runtime.__toESM(require("@langchain/core/prompts"));
-const __langchain_core_output_parsers = require_rolldown_runtime.__toESM(require("@langchain/core/output_parsers"));
-const __langchain_core_retrievers = require_rolldown_runtime.__toESM(require("@langchain/core/retrievers"));
+let _langchain_core_prompts = require("@langchain/core/prompts");
+let _langchain_core_output_parsers = require("@langchain/core/output_parsers");
+let _langchain_core_retrievers = require("@langchain/core/retrievers");
 
 //#region src/retrievers/multi_query.ts
-var multi_query_exports = {};
-require_rolldown_runtime.__export(multi_query_exports, { MultiQueryRetriever: () => MultiQueryRetriever });
-var LineListOutputParser = class extends __langchain_core_output_parsers.BaseOutputParser {
+var multi_query_exports = /* @__PURE__ */ require_runtime.__exportAll({ MultiQueryRetriever: () => MultiQueryRetriever });
+var LineListOutputParser = class extends _langchain_core_output_parsers.BaseOutputParser {
 	static lc_name() {
 		return "LineListOutputParser";
 	}
@@ -21,14 +21,13 @@ var LineListOutputParser = class extends __langchain_core_output_parsers.BaseOut
 		const endKeyIndex = text.indexOf("</questions>");
 		const questionsStartIndex = startKeyIndex === -1 ? 0 : startKeyIndex + 11;
 		const questionsEndIndex = endKeyIndex === -1 ? text.length : endKeyIndex;
-		const lines = text.slice(questionsStartIndex, questionsEndIndex).trim().split("\n").filter((line) => line.trim() !== "");
-		return { lines };
+		return { lines: text.slice(questionsStartIndex, questionsEndIndex).trim().split("\n").filter((line) => line.trim() !== "") };
 	}
 	getFormatInstructions() {
 		throw new Error("Not implemented.");
 	}
 };
-const DEFAULT_QUERY_PROMPT = /* @__PURE__ */ new __langchain_core_prompts.PromptTemplate({
+const DEFAULT_QUERY_PROMPT = /* @__PURE__ */ new _langchain_core_prompts.PromptTemplate({
 	inputVariables: ["question", "queryCount"],
 	template: `You are an AI language model assistant. Your task is
 to generate {queryCount} different versions of the given user
@@ -60,7 +59,7 @@ Original question: {question}`
 * );
 * ```
 */
-var MultiQueryRetriever = class extends __langchain_core_retrievers.BaseRetriever {
+var MultiQueryRetriever = class extends _langchain_core_retrievers.BaseRetriever {
 	static lc_name() {
 		return "MultiQueryRetriever";
 	}
@@ -85,12 +84,11 @@ var MultiQueryRetriever = class extends __langchain_core_retrievers.BaseRetrieve
 		this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;
 	}
 	static fromLLM(fields) {
-		const { retriever, llm, prompt = DEFAULT_QUERY_PROMPT, queryCount, parserKey,...rest } = fields;
-		const outputParser = new LineListOutputParser();
+		const { retriever, llm, prompt = DEFAULT_QUERY_PROMPT, queryCount, parserKey, ...rest } = fields;
 		const llmChain = new require_llm_chain.LLMChain({
 			llm,
 			prompt,
-			outputParser
+			outputParser: new LineListOutputParser()
 		});
 		return new this({
 			retriever,
@@ -101,11 +99,10 @@ var MultiQueryRetriever = class extends __langchain_core_retrievers.BaseRetrieve
 		});
 	}
 	async _generateQueries(question, runManager) {
-		const response = await this.llmChain.call({
+		const lines = (await this.llmChain.call({
 			question,
 			queryCount: this.queryCount
-		}, runManager?.getChild());
-		const lines = response.text[this.parserKey] || [];
+		}, runManager?.getChild())).text[this.parserKey] || [];
 		if (this.verbose) console.log(`Generated queries: ${lines}`);
 		return lines;
 	}
@@ -123,8 +120,7 @@ var MultiQueryRetriever = class extends __langchain_core_retrievers.BaseRetrieve
 			const key = `${doc.pageContent}:${JSON.stringify(Object.entries(doc.metadata).sort())}`;
 			uniqueDocumentsDict[key] = doc;
 		}
-		const uniqueDocuments = Object.values(uniqueDocumentsDict);
-		return uniqueDocuments;
+		return Object.values(uniqueDocumentsDict);
 	}
 	async _getRelevantDocuments(question, runManager) {
 		const queries = await this._generateQueries(question, runManager);

package/dist/retrievers/multi_query.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"multi_query.cjs","names":["BaseOutputParser","text: string","PromptTemplate","BaseRetriever","fields: MultiQueryRetrieverInput","fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n      llm: BaseLanguageModelInterface;\n      prompt?: BasePromptTemplate;\n    }","LLMChain","question: string","runManager?: CallbackManagerForRetrieverRun","queries: string[]","documents: Document[]","uniqueDocumentsDict: { [key: string]: Document }"],"sources":["../../src/retrievers/multi_query.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport {\n  BaseRetriever,\n  type BaseRetrieverInput,\n  type BaseRetrieverInterface,\n} from \"@langchain/core/retrievers\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseOutputParser } from \"@langchain/core/output_parsers\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../chains/llm_chain.js\";\nimport type { BaseDocumentCompressor } from \"./document_compressors/index.js\";\n\ninterface LineList {\n  lines: string[];\n}\n\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\nexport type MultiDocs = Document<Record<string, any>>[];\n\nclass LineListOutputParser extends BaseOutputParser<LineList> {\n  static lc_name() {\n    return \"LineListOutputParser\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  async parse(text: string): Promise<LineList> {\n    const startKeyIndex = text.indexOf(\"<questions>\");\n    const endKeyIndex = text.indexOf(\"</questions>\");\n    const questionsStartIndex =\n      startKeyIndex === -1 ? 0 : startKeyIndex + \"<questions>\".length;\n    const questionsEndIndex = endKeyIndex === -1 ? text.length : endKeyIndex;\n    const lines = text\n      .slice(questionsStartIndex, questionsEndIndex)\n      .trim()\n      .split(\"\\n\")\n      .filter((line) => line.trim() !== \"\");\n    return { lines };\n  }\n\n  getFormatInstructions(): string {\n    throw new Error(\"Not implemented.\");\n  }\n}\n\n// Create template\nconst DEFAULT_QUERY_PROMPT = /* #__PURE__ */ new PromptTemplate({\n  inputVariables: [\"question\", \"queryCount\"],\n  template: `You are an AI language model assistant. Your task is\nto generate {queryCount} different versions of the given user\nquestion to retrieve relevant documents from a vector database.\nBy generating multiple perspectives on the user question,\nyour goal is to help the user overcome some of the limitations\nof distance-based similarity search.\n\nProvide these alternative questions separated by newlines between XML tags. For example:\n\n<questions>\nQuestion 1\nQuestion 2\nQuestion 3\n</questions>\n\nOriginal question: {question}`,\n});\n\nexport interface MultiQueryRetrieverInput extends BaseRetrieverInput {\n  retriever: BaseRetrieverInterface;\n  /** @deprecated Pass a custom prompt into `.fromLLM` instead. */\n  llmChain: LLMChain<LineList>;\n  queryCount?: number;\n  parserKey?: string;\n  documentCompressor?: BaseDocumentCompressor | undefined;\n  documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;\n}\n\n/**\n * @example\n * ```typescript\n * const retriever = new MultiQueryRetriever.fromLLM({\n *   llm: new ChatAnthropic({}),\n *   retriever: new MemoryVectorStore().asRetriever(),\n *   verbose: true,\n * });\n * const retrievedDocs = await retriever.invoke(\n *   \"What are mitochondria made of?\",\n * );\n * ```\n */\nexport class MultiQueryRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"MultiQueryRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  private retriever: BaseRetrieverInterface;\n\n  private llmChain: LLMChain<LineList>;\n\n  private queryCount = 3;\n\n  private parserKey = \"lines\";\n\n  documentCompressor: BaseDocumentCompressor | undefined;\n\n  documentCompressorFilteringFn?: MultiQueryRetrieverInput[\"documentCompressorFilteringFn\"];\n\n  constructor(fields: MultiQueryRetrieverInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llmChain = fields.llmChain;\n    this.queryCount = fields.queryCount ?? this.queryCount;\n    this.parserKey = fields.parserKey ?? this.parserKey;\n    this.documentCompressor = fields.documentCompressor;\n    this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;\n  }\n\n  static fromLLM(\n    fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n      llm: BaseLanguageModelInterface;\n      prompt?: BasePromptTemplate;\n    }\n  ): MultiQueryRetriever {\n    const {\n      retriever,\n      llm,\n      prompt = DEFAULT_QUERY_PROMPT,\n      queryCount,\n      parserKey,\n      ...rest\n    } = fields;\n    const outputParser = new LineListOutputParser();\n    const llmChain = new LLMChain({ llm, prompt, outputParser });\n    return new this({ retriever, llmChain, queryCount, parserKey, ...rest });\n  }\n\n  // Generate the different queries for each retrieval, using our llmChain\n  private async _generateQueries(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<string[]> {\n    const response = await this.llmChain.call(\n      { question, queryCount: this.queryCount },\n      runManager?.getChild()\n    );\n    const lines = response.text[this.parserKey] || [];\n    if (this.verbose) {\n      console.log(`Generated queries: ${lines}`);\n    }\n    return lines;\n  }\n\n  // Retrieve documents using the original retriever\n  private async _retrieveDocuments(\n    queries: string[],\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const documents: Document[] = [];\n    await Promise.all(\n      queries.map(async (query) => {\n        const docs = await this.retriever.invoke(query, runManager?.getChild());\n        documents.push(...docs);\n      })\n    );\n    return documents;\n  }\n\n  // Deduplicate the documents that were returned in multiple retrievals\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const uniqueDocumentsDict: { [key: string]: Document } = {};\n\n    for (const doc of documents) {\n      const key = `${doc.pageContent}:${JSON.stringify(\n        Object.entries(doc.metadata).sort()\n      )}`;\n      uniqueDocumentsDict[key] = doc;\n    }\n\n    const uniqueDocuments = Object.values(uniqueDocumentsDict);\n    return uniqueDocuments;\n  }\n\n  async _getRelevantDocuments(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const queries = await this._generateQueries(question, runManager);\n    const documents = await this._retrieveDocuments(queries, runManager);\n    const uniqueDocuments = this._uniqueUnion(documents);\n\n    let outputDocs = uniqueDocuments;\n    if (this.documentCompressor && uniqueDocuments.length) {\n      outputDocs = await this.documentCompressor.compressDocuments(\n        uniqueDocuments,\n        question,\n        runManager?.getChild()\n      );\n      if (this.documentCompressorFilteringFn) {\n        outputDocs = this.documentCompressorFilteringFn(outputDocs);\n      }\n    }\n\n    return outputDocs;\n  }\n}\n"],"mappings":";;;;;;;;;AAoBA,IAAM,uBAAN,cAAmCA,iDAA2B;CAC5D,OAAO,UAAU;AACf,SAAO;CACR;CAED,eAAe;EAAC;EAAa;EAAc;CAAa;CAExD,MAAM,MAAMC,MAAiC;EAC3C,MAAM,gBAAgB,KAAK,QAAQ,cAAc;EACjD,MAAM,cAAc,KAAK,QAAQ,eAAe;EAChD,MAAM,sBACJ,kBAAkB,KAAK,IAAI,gBAAgB;EAC7C,MAAM,oBAAoB,gBAAgB,KAAK,KAAK,SAAS;EAC7D,MAAM,QAAQ,KACX,MAAM,qBAAqB,kBAAkB,CAC7C,MAAM,CACN,MAAM,KAAK,CACX,OAAO,CAAC,SAAS,KAAK,MAAM,KAAK,GAAG;AACvC,SAAO,EAAE,MAAO;CACjB;CAED,wBAAgC;AAC9B,QAAM,IAAI,MAAM;CACjB;AACF;AAGD,MAAM,uCAAuC,IAAIC,wCAAe;CAC9D,gBAAgB,CAAC,YAAY,YAAa;CAC1C,UAAU,CAAC;;;;;;;;;;;;;;;6BAegB,CAAC;AAC7B;;;;;;;;;;;;;;AAyBD,IAAa,sBAAb,cAAyCC,0CAAc;CACrD,OAAO,UAAU;AACf,SAAO;CACR;CAED,eAAe;EAAC;EAAa;EAAc;CAAa;CAExD,AAAQ;CAER,AAAQ;CAER,AAAQ,aAAa;CAErB,AAAQ,YAAY;CAEpB;CAEA;CAEA,YAAYC,QAAkC;EAC5C,MAAM,OAAO;EACb,KAAK,YAAY,OAAO;EACxB,KAAK,WAAW,OAAO;EACvB,KAAK,aAAa,OAAO,cAAc,KAAK;EAC5C,KAAK,YAAY,OAAO,aAAa,KAAK;EAC1C,KAAK,qBAAqB,OAAO;EACjC,KAAK,gCAAgC,OAAO;CAC7C;CAED,OAAO,QACLC,QAIqB;EACrB,MAAM,EACJ,WACA,KACA,SAAS,sBACT,YACA,UACA,GAAG,MACJ,GAAG;EACJ,MAAM,eAAe,IAAI;EACzB,MAAM,WAAW,IAAIC,2BAAS;GAAE;GAAK;GAAQ;EAAc;AAC3D,SAAO,IAAI,KAAK;GAAE;GAAW;GAAU;GAAY;GAAW,GAAG;EAAM;CACxE;CAGD,MAAc,iBACZC,UACAC,YACmB;EACnB,MAAM,WAAW,MAAM,KAAK,SAAS,KACnC;GAAE;GAAU,YAAY,KAAK;EAAY,GACzC,YAAY,UAAU,CACvB;EACD,MAAM,QAAQ,SAAS,KAAK,KAAK,cAAc,CAAE;AACjD,MAAI,KAAK,SACP,QAAQ,IAAI,CAAC,mBAAmB,EAAE,OAAO,CAAC;AAE5C,SAAO;CACR;CAGD,MAAc,mBACZC,SACAD,YACqB;EACrB,MAAME,YAAwB,CAAE;EAChC,MAAM,QAAQ,IACZ,QAAQ,IAAI,OAAO,UAAU;GAC3B,MAAM,OAAO,MAAM,KAAK,UAAU,OAAO,OAAO,YAAY,UAAU,CAAC;GACvE,UAAU,KAAK,GAAG,KAAK;EACxB,EAAC,CACH;AACD,SAAO;CACR;CAGD,AAAQ,aAAaA,WAAmC;EACtD,MAAMC,sBAAmD,CAAE;AAE3D,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,EAAE,KAAK,UACrC,OAAO,QAAQ,IAAI,SAAS,CAAC,MAAM,CACpC,EAAE;GACH,oBAAoB,OAAO;EAC5B;EAED,MAAM,kBAAkB,OAAO,OAAO,oBAAoB;AAC1D,SAAO;CACR;CAED,MAAM,sBACJJ,UACAC,YACqB;EACrB,MAAM,UAAU,MAAM,KAAK,iBAAiB,UAAU,WAAW;EACjE,MAAM,YAAY,MAAM,KAAK,mBAAmB,SAAS,WAAW;EACpE,MAAM,kBAAkB,KAAK,aAAa,UAAU;EAEpD,IAAI,aAAa;AACjB,MAAI,KAAK,sBAAsB,gBAAgB,QAAQ;GACrD,aAAa,MAAM,KAAK,mBAAmB,kBACzC,iBACA,UACA,YAAY,UAAU,CACvB;AACD,OAAI,KAAK,+BACP,aAAa,KAAK,8BAA8B,WAAW;EAE9D;AAED,SAAO;CACR;AACF"}
+{"version":3,"file":"multi_query.cjs","names":["BaseOutputParser","PromptTemplate","BaseRetriever","LLMChain"],"sources":["../../src/retrievers/multi_query.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport {\n  BaseRetriever,\n  type BaseRetrieverInput,\n  type BaseRetrieverInterface,\n} from \"@langchain/core/retrievers\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseOutputParser } from \"@langchain/core/output_parsers\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../chains/llm_chain.js\";\nimport type { BaseDocumentCompressor } from \"./document_compressors/index.js\";\n\ninterface LineList {\n  lines: string[];\n}\n\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\nexport type MultiDocs = Document<Record<string, any>>[];\n\nclass LineListOutputParser extends BaseOutputParser<LineList> {\n  static lc_name() {\n    return \"LineListOutputParser\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  async parse(text: string): Promise<LineList> {\n    const startKeyIndex = text.indexOf(\"<questions>\");\n    const endKeyIndex = text.indexOf(\"</questions>\");\n    const questionsStartIndex =\n      startKeyIndex === -1 ? 0 : startKeyIndex + \"<questions>\".length;\n    const questionsEndIndex = endKeyIndex === -1 ? text.length : endKeyIndex;\n    const lines = text\n      .slice(questionsStartIndex, questionsEndIndex)\n      .trim()\n      .split(\"\\n\")\n      .filter((line) => line.trim() !== \"\");\n    return { lines };\n  }\n\n  getFormatInstructions(): string {\n    throw new Error(\"Not implemented.\");\n  }\n}\n\n// Create template\nconst DEFAULT_QUERY_PROMPT = /* #__PURE__ */ new PromptTemplate({\n  inputVariables: [\"question\", \"queryCount\"],\n  template: `You are an AI language model assistant. Your task is\nto generate {queryCount} different versions of the given user\nquestion to retrieve relevant documents from a vector database.\nBy generating multiple perspectives on the user question,\nyour goal is to help the user overcome some of the limitations\nof distance-based similarity search.\n\nProvide these alternative questions separated by newlines between XML tags. For example:\n\n<questions>\nQuestion 1\nQuestion 2\nQuestion 3\n</questions>\n\nOriginal question: {question}`,\n});\n\nexport interface MultiQueryRetrieverInput extends BaseRetrieverInput {\n  retriever: BaseRetrieverInterface;\n  /** @deprecated Pass a custom prompt into `.fromLLM` instead. */\n  llmChain: LLMChain<LineList>;\n  queryCount?: number;\n  parserKey?: string;\n  documentCompressor?: BaseDocumentCompressor | undefined;\n  documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;\n}\n\n/**\n * @example\n * ```typescript\n * const retriever = new MultiQueryRetriever.fromLLM({\n *   llm: new ChatAnthropic({}),\n *   retriever: new MemoryVectorStore().asRetriever(),\n *   verbose: true,\n * });\n * const retrievedDocs = await retriever.invoke(\n *   \"What are mitochondria made of?\",\n * );\n * ```\n */\nexport class MultiQueryRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"MultiQueryRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  private retriever: BaseRetrieverInterface;\n\n  private llmChain: LLMChain<LineList>;\n\n  private queryCount = 3;\n\n  private parserKey = \"lines\";\n\n  documentCompressor: BaseDocumentCompressor | undefined;\n\n  documentCompressorFilteringFn?: MultiQueryRetrieverInput[\"documentCompressorFilteringFn\"];\n\n  constructor(fields: MultiQueryRetrieverInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llmChain = fields.llmChain;\n    this.queryCount = fields.queryCount ?? this.queryCount;\n    this.parserKey = fields.parserKey ?? this.parserKey;\n    this.documentCompressor = fields.documentCompressor;\n    this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;\n  }\n\n  static fromLLM(\n    fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n      llm: BaseLanguageModelInterface;\n      prompt?: BasePromptTemplate;\n    }\n  ): MultiQueryRetriever {\n    const {\n      retriever,\n      llm,\n      prompt = DEFAULT_QUERY_PROMPT,\n      queryCount,\n      parserKey,\n      ...rest\n    } = fields;\n    const outputParser = new LineListOutputParser();\n    const llmChain = new LLMChain({ llm, prompt, outputParser });\n    return new this({ retriever, llmChain, queryCount, parserKey, ...rest });\n  }\n\n  // Generate the different queries for each retrieval, using our llmChain\n  private async _generateQueries(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<string[]> {\n    const response = await this.llmChain.call(\n      { question, queryCount: this.queryCount },\n      runManager?.getChild()\n    );\n    const lines = response.text[this.parserKey] || [];\n    if (this.verbose) {\n      console.log(`Generated queries: ${lines}`);\n    }\n    return lines;\n  }\n\n  // Retrieve documents using the original retriever\n  private async _retrieveDocuments(\n    queries: string[],\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const documents: Document[] = [];\n    await Promise.all(\n      queries.map(async (query) => {\n        const docs = await this.retriever.invoke(query, runManager?.getChild());\n        documents.push(...docs);\n      })\n    );\n    return documents;\n  }\n\n  // Deduplicate the documents that were returned in multiple retrievals\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const uniqueDocumentsDict: { [key: string]: Document } = {};\n\n    for (const doc of documents) {\n      const key = `${doc.pageContent}:${JSON.stringify(\n        Object.entries(doc.metadata).sort()\n      )}`;\n      uniqueDocumentsDict[key] = doc;\n    }\n\n    const uniqueDocuments = Object.values(uniqueDocumentsDict);\n    return uniqueDocuments;\n  }\n\n  async _getRelevantDocuments(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const queries = await this._generateQueries(question, runManager);\n    const documents = await this._retrieveDocuments(queries, runManager);\n    const uniqueDocuments = this._uniqueUnion(documents);\n\n    let outputDocs = uniqueDocuments;\n    if (this.documentCompressor && uniqueDocuments.length) {\n      outputDocs = await this.documentCompressor.compressDocuments(\n        uniqueDocuments,\n        question,\n        runManager?.getChild()\n      );\n      if (this.documentCompressorFilteringFn) {\n        outputDocs = this.documentCompressorFilteringFn(outputDocs);\n      }\n    }\n\n    return outputDocs;\n  }\n}\n"],"mappings":";;;;;;;;;AAoBA,IAAM,uBAAN,cAAmCA,gDAA2B;CAC5D,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAa;CAExD,MAAM,MAAM,MAAiC;EAC3C,MAAM,gBAAgB,KAAK,QAAQ,cAAc;EACjD,MAAM,cAAc,KAAK,QAAQ,eAAe;EAChD,MAAM,sBACJ,kBAAkB,KAAK,IAAI,gBAAgB;EAC7C,MAAM,oBAAoB,gBAAgB,KAAK,KAAK,SAAS;AAM7D,SAAO,EAAE,OALK,KACX,MAAM,qBAAqB,kBAAkB,CAC7C,MAAM,CACN,MAAM,KAAK,CACX,QAAQ,SAAS,KAAK,MAAM,KAAK,GAAG,EACvB;;CAGlB,wBAAgC;AAC9B,QAAM,IAAI,MAAM,mBAAmB;;;AAKvC,MAAM,uCAAuC,IAAIC,uCAAe;CAC9D,gBAAgB,CAAC,YAAY,aAAa;CAC1C,UAAU;;;;;;;;;;;;;;;;CAgBX,CAAC;;;;;;;;;;;;;;AAyBF,IAAa,sBAAb,cAAyCC,yCAAc;CACrD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAa;CAExD,AAAQ;CAER,AAAQ;CAER,AAAQ,aAAa;CAErB,AAAQ,YAAY;CAEpB;CAEA;CAEA,YAAY,QAAkC;AAC5C,QAAM,OAAO;AACb,OAAK,YAAY,OAAO;AACxB,OAAK,WAAW,OAAO;AACvB,OAAK,aAAa,OAAO,cAAc,KAAK;AAC5C,OAAK,YAAY,OAAO,aAAa,KAAK;AAC1C,OAAK,qBAAqB,OAAO;AACjC,OAAK,gCAAgC,OAAO;;CAG9C,OAAO,QACL,QAIqB;EACrB,MAAM,EACJ,WACA,KACA,SAAS,sBACT,YACA,WACA,GAAG,SACD;EAEJ,MAAM,WAAW,IAAIC,2BAAS;GAAE;GAAK;GAAQ,cADxB,IAAI,sBAAsB;GACY,CAAC;AAC5D,SAAO,IAAI,KAAK;GAAE;GAAW;GAAU;GAAY;GAAW,GAAG;GAAM,CAAC;;CAI1E,MAAc,iBACZ,UACA,YACmB;EAKnB,MAAM,SAJW,MAAM,KAAK,SAAS,KACnC;GAAE;GAAU,YAAY,KAAK;GAAY,EACzC,YAAY,UAAU,CACvB,EACsB,KAAK,KAAK,cAAc,EAAE;AACjD,MAAI,KAAK,QACP,SAAQ,IAAI,sBAAsB,QAAQ;AAE5C,SAAO;;CAIT,MAAc,mBACZ,SACA,YACqB;EACrB,MAAM,YAAwB,EAAE;AAChC,QAAM,QAAQ,IACZ,QAAQ,IAAI,OAAO,UAAU;GAC3B,MAAM,OAAO,MAAM,KAAK,UAAU,OAAO,OAAO,YAAY,UAAU,CAAC;AACvE,aAAU,KAAK,GAAG,KAAK;IACvB,CACH;AACD,SAAO;;CAIT,AAAQ,aAAa,WAAmC;EACtD,MAAM,sBAAmD,EAAE;AAE3D,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,GAAG,IAAI,YAAY,GAAG,KAAK,UACrC,OAAO,QAAQ,IAAI,SAAS,CAAC,MAAM,CACpC;AACD,uBAAoB,OAAO;;AAI7B,SADwB,OAAO,OAAO,oBAAoB;;CAI5D,MAAM,sBACJ,UACA,YACqB;EACrB,MAAM,UAAU,MAAM,KAAK,iBAAiB,UAAU,WAAW;EACjE,MAAM,YAAY,MAAM,KAAK,mBAAmB,SAAS,WAAW;EACpE,MAAM,kBAAkB,KAAK,aAAa,UAAU;EAEpD,IAAI,aAAa;AACjB,MAAI,KAAK,sBAAsB,gBAAgB,QAAQ;AACrD,gBAAa,MAAM,KAAK,mBAAmB,kBACzC,iBACA,UACA,YAAY,UAAU,CACvB;AACD,OAAI,KAAK,8BACP,cAAa,KAAK,8BAA8B,WAAW;;AAI/D,SAAO"}

package/dist/retrievers/multi_query.d.cts.map

@@ -1 +1 @@
-{"version":3,"file":"multi_query.d.cts","names":["BaseLanguageModelInterface","BaseRetriever","BaseRetrieverInput","BaseRetrieverInterface","Document","BasePromptTemplate","CallbackManagerForRetrieverRun","LLMChain","BaseDocumentCompressor","LineList","MultiDocs","Record","MultiQueryRetrieverInput","MultiQueryRetriever","Omit","Promise"],"sources":["../../src/retrievers/multi_query.d.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { BaseRetriever, type BaseRetrieverInput, type BaseRetrieverInterface } from \"@langchain/core/retrievers\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BasePromptTemplate } from \"@langchain/core/prompts\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../chains/llm_chain.js\";\nimport type { BaseDocumentCompressor } from \"./document_compressors/index.js\";\ninterface LineList {\n    lines: string[];\n}\nexport type MultiDocs = Document<Record<string, any>>[];\nexport interface MultiQueryRetrieverInput extends BaseRetrieverInput {\n    retriever: BaseRetrieverInterface;\n    /** @deprecated Pass a custom prompt into `.fromLLM` instead. */\n    llmChain: LLMChain<LineList>;\n    queryCount?: number;\n    parserKey?: string;\n    documentCompressor?: BaseDocumentCompressor | undefined;\n    documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;\n}\n/**\n * @example\n * ```typescript\n * const retriever = new MultiQueryRetriever.fromLLM({\n *   llm: new ChatAnthropic({}),\n *   retriever: new MemoryVectorStore().asRetriever(),\n *   verbose: true,\n * });\n * const retrievedDocs = await retriever.invoke(\n *   \"What are mitochondria made of?\",\n * );\n * ```\n */\nexport declare class MultiQueryRetriever extends BaseRetriever {\n    static lc_name(): string;\n    lc_namespace: string[];\n    private retriever;\n    private llmChain;\n    private queryCount;\n    private parserKey;\n    documentCompressor: BaseDocumentCompressor | undefined;\n    documentCompressorFilteringFn?: MultiQueryRetrieverInput[\"documentCompressorFilteringFn\"];\n    constructor(fields: MultiQueryRetrieverInput);\n    static fromLLM(fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n        llm: BaseLanguageModelInterface;\n        prompt?: BasePromptTemplate;\n    }): MultiQueryRetriever;\n    private _generateQueries;\n    private _retrieveDocuments;\n    private _uniqueUnion;\n    _getRelevantDocuments(question: string, runManager?: CallbackManagerForRetrieverRun): Promise<Document[]>;\n}\nexport {};\n//# sourceMappingURL=multi_query.d.ts.map"],"mappings":";;;;;;;;;UAOUS,QAAAA;;AADoE;AAIlEC,KAAAA,SAAAA,GAAYN,QAASO,CAAAA,MAATP,CAAAA,MAAQ,EAAA,GAAA,CAAA,CAAA,EAAA;AACfQ,UAAAA,wBAAAA,SAAiCV,kBAAT,CAAA;EAC1BC,SAAAA,EAAAA,sBAAAA;EAEQM;EAATF,QAAAA,EAAAA,QAAAA,CAASE,QAATF,CAAAA;EAGWC,UAAAA,CAAAA,EAAAA,MAAAA;EACkBE,SAAAA,CAAAA,EAAAA,MAAAA;EAAcA,kBAAAA,CAAAA,EADhCF,sBACgCE,GAAAA,SAAAA;EAPPR,6BAAAA,CAAAA,EAAAA,CAAAA,IAAAA,EAOPQ,SAPOR,EAAAA,GAOOQ,SAPPR;AAAkB;AAsBpE;;;;;;;;;;;;;AAA8D,cAAzCW,mBAAAA,SAA4BZ,aAAAA,CAAa;;;;;;;sBAOtCO;kCACYI;sBACZA;yBACGE,KAAKF;SACnBZ;aACIK;MACTQ;;;;uDAIiDP,iCAAiCS,QAAQX"}
+{"version":3,"file":"multi_query.d.cts","names":[],"sources":["../../src/retrievers/multi_query.ts"],"mappings":";;;;;;;;;UAaU,QAAA;EACR,KAAA;AAAA;AAAA,KAIU,SAAA,GAAY,QAAA,CAAS,MAAA;AAAA,UAiDhB,wBAAA,SAAiC,kBAAA;EAChD,SAAA,EAAW,sBAAA;EAtDN;EAwDL,QAAA,EAAU,QAAA,CAAS,QAAA;EACnB,UAAA;EACA,SAAA;EACA,kBAAA,GAAqB,sBAAA;EACrB,6BAAA,IAAiC,IAAA,EAAM,SAAA,KAAc,SAAA;AAAA;;;;;;;;;;;;;;cAgB1C,mBAAA,SAA4B,aAAA;EAAA,OAChC,OAAA,CAAA;EAIP,YAAA;EAAA,QAEQ,SAAA;EAAA,QAEA,QAAA;EAAA,QAEA,UAAA;EAAA,QAEA,SAAA;EAER,kBAAA,EAAoB,sBAAA;EAEpB,6BAAA,GAAgC,wBAAA;EAEhC,WAAA,CAAY,MAAA,EAAQ,wBAAA;EAAA,OAUb,OAAA,CACL,MAAA,EAAQ,IAAA,CAAK,wBAAA;IACX,GAAA,EAAK,0BAAA;IACL,MAAA,GAAS,kBAAA;EAAA,IAEV,mBAAA;EAAA,QAeW,gBAAA;EAAA,QAgBA,kBAAA;EAAA,QAeN,YAAA;EAcF,qBAAA,CACJ,QAAA,UACA,UAAA,GAAa,8BAAA,GACZ,OAAA,CAAQ,QAAA;AAAA"}

package/dist/retrievers/multi_query.d.ts

@@ -1,9 +1,9 @@
 import { LLMChain } from "../chains/llm_chain.js";
 import { BaseDocumentCompressor } from "./document_compressors/index.js";
-import { Document } from "@langchain/core/documents";
 import { BasePromptTemplate } from "@langchain/core/prompts";
 import { BaseLanguageModelInterface } from "@langchain/core/language_models/base";
 import { CallbackManagerForRetrieverRun } from "@langchain/core/callbacks/manager";
+import { Document } from "@langchain/core/documents";
 import { BaseRetriever, BaseRetrieverInput, BaseRetrieverInterface } from "@langchain/core/retrievers";
 
 //#region src/retrievers/multi_query.d.ts

package/dist/retrievers/multi_query.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"multi_query.d.ts","names":["BaseLanguageModelInterface","BaseRetriever","BaseRetrieverInput","BaseRetrieverInterface","Document","BasePromptTemplate","CallbackManagerForRetrieverRun","LLMChain","BaseDocumentCompressor","LineList","MultiDocs","Record","MultiQueryRetrieverInput","MultiQueryRetriever","Omit","Promise"],"sources":["../../src/retrievers/multi_query.d.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { BaseRetriever, type BaseRetrieverInput, type BaseRetrieverInterface } from \"@langchain/core/retrievers\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BasePromptTemplate } from \"@langchain/core/prompts\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../chains/llm_chain.js\";\nimport type { BaseDocumentCompressor } from \"./document_compressors/index.js\";\ninterface LineList {\n    lines: string[];\n}\nexport type MultiDocs = Document<Record<string, any>>[];\nexport interface MultiQueryRetrieverInput extends BaseRetrieverInput {\n    retriever: BaseRetrieverInterface;\n    /** @deprecated Pass a custom prompt into `.fromLLM` instead. */\n    llmChain: LLMChain<LineList>;\n    queryCount?: number;\n    parserKey?: string;\n    documentCompressor?: BaseDocumentCompressor | undefined;\n    documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;\n}\n/**\n * @example\n * ```typescript\n * const retriever = new MultiQueryRetriever.fromLLM({\n *   llm: new ChatAnthropic({}),\n *   retriever: new MemoryVectorStore().asRetriever(),\n *   verbose: true,\n * });\n * const retrievedDocs = await retriever.invoke(\n *   \"What are mitochondria made of?\",\n * );\n * ```\n */\nexport declare class MultiQueryRetriever extends BaseRetriever {\n    static lc_name(): string;\n    lc_namespace: string[];\n    private retriever;\n    private llmChain;\n    private queryCount;\n    private parserKey;\n    documentCompressor: BaseDocumentCompressor | undefined;\n    documentCompressorFilteringFn?: MultiQueryRetrieverInput[\"documentCompressorFilteringFn\"];\n    constructor(fields: MultiQueryRetrieverInput);\n    static fromLLM(fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n        llm: BaseLanguageModelInterface;\n        prompt?: BasePromptTemplate;\n    }): MultiQueryRetriever;\n    private _generateQueries;\n    private _retrieveDocuments;\n    private _uniqueUnion;\n    _getRelevantDocuments(question: string, runManager?: CallbackManagerForRetrieverRun): Promise<Document[]>;\n}\nexport {};\n//# sourceMappingURL=multi_query.d.ts.map"],"mappings":";;;;;;;;;UAOUS,QAAAA;;AADoE;AAIlEC,KAAAA,SAAAA,GAAYN,QAASO,CAAAA,MAATP,CAAAA,MAAQ,EAAA,GAAA,CAAA,CAAA,EAAA;AACfQ,UAAAA,wBAAAA,SAAiCV,kBAAT,CAAA;EAC1BC,SAAAA,EAAAA,sBAAAA;EAEQM;EAATF,QAAAA,EAAAA,QAAAA,CAASE,QAATF,CAAAA;EAGWC,UAAAA,CAAAA,EAAAA,MAAAA;EACkBE,SAAAA,CAAAA,EAAAA,MAAAA;EAAcA,kBAAAA,CAAAA,EADhCF,sBACgCE,GAAAA,SAAAA;EAPPR,6BAAAA,CAAAA,EAAAA,CAAAA,IAAAA,EAOPQ,SAPOR,EAAAA,GAOOQ,SAPPR;AAAkB;AAsBpE;;;;;;;;;;;;;AAA8D,cAAzCW,mBAAAA,SAA4BZ,aAAAA,CAAa;;;;;;;sBAOtCO;kCACYI;sBACZA;yBACGE,KAAKF;SACnBZ;aACIK;MACTQ;;;;uDAIiDP,iCAAiCS,QAAQX"}
+{"version":3,"file":"multi_query.d.ts","names":[],"sources":["../../src/retrievers/multi_query.ts"],"mappings":";;;;;;;;;UAaU,QAAA;EACR,KAAA;AAAA;AAAA,KAIU,SAAA,GAAY,QAAA,CAAS,MAAA;AAAA,UAiDhB,wBAAA,SAAiC,kBAAA;EAChD,SAAA,EAAW,sBAAA;EAtDN;EAwDL,QAAA,EAAU,QAAA,CAAS,QAAA;EACnB,UAAA;EACA,SAAA;EACA,kBAAA,GAAqB,sBAAA;EACrB,6BAAA,IAAiC,IAAA,EAAM,SAAA,KAAc,SAAA;AAAA;;;;;;;;;;;;;;cAgB1C,mBAAA,SAA4B,aAAA;EAAA,OAChC,OAAA,CAAA;EAIP,YAAA;EAAA,QAEQ,SAAA;EAAA,QAEA,QAAA;EAAA,QAEA,UAAA;EAAA,QAEA,SAAA;EAER,kBAAA,EAAoB,sBAAA;EAEpB,6BAAA,GAAgC,wBAAA;EAEhC,WAAA,CAAY,MAAA,EAAQ,wBAAA;EAAA,OAUb,OAAA,CACL,MAAA,EAAQ,IAAA,CAAK,wBAAA;IACX,GAAA,EAAK,0BAAA;IACL,MAAA,GAAS,kBAAA;EAAA,IAEV,mBAAA;EAAA,QAeW,gBAAA;EAAA,QAgBA,kBAAA;EAAA,QAeN,YAAA;EAcF,qBAAA,CACJ,QAAA,UACA,UAAA,GAAa,8BAAA,GACZ,OAAA,CAAQ,QAAA;AAAA"}

package/dist/retrievers/multi_query.js

@@ -1,12 +1,11 @@
-import { __export } from "../_virtual/rolldown_runtime.js";
+import { __exportAll } from "../_virtual/_rolldown/runtime.js";
 import { LLMChain } from "../chains/llm_chain.js";
 import { PromptTemplate } from "@langchain/core/prompts";
 import { BaseOutputParser } from "@langchain/core/output_parsers";
 import { BaseRetriever } from "@langchain/core/retrievers";
 
 //#region src/retrievers/multi_query.ts
-var multi_query_exports = {};
-__export(multi_query_exports, { MultiQueryRetriever: () => MultiQueryRetriever });
+var multi_query_exports = /* @__PURE__ */ __exportAll({ MultiQueryRetriever: () => MultiQueryRetriever });
 var LineListOutputParser = class extends BaseOutputParser {
 	static lc_name() {
 		return "LineListOutputParser";
@@ -21,8 +20,7 @@ var LineListOutputParser = class extends BaseOutputParser {
 		const endKeyIndex = text.indexOf("</questions>");
 		const questionsStartIndex = startKeyIndex === -1 ? 0 : startKeyIndex + 11;
 		const questionsEndIndex = endKeyIndex === -1 ? text.length : endKeyIndex;
-		const lines = text.slice(questionsStartIndex, questionsEndIndex).trim().split("\n").filter((line) => line.trim() !== "");
-		return { lines };
+		return { lines: text.slice(questionsStartIndex, questionsEndIndex).trim().split("\n").filter((line) => line.trim() !== "") };
 	}
 	getFormatInstructions() {
 		throw new Error("Not implemented.");
@@ -85,12 +83,11 @@ var MultiQueryRetriever = class extends BaseRetriever {
 		this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;
 	}
 	static fromLLM(fields) {
-		const { retriever, llm, prompt = DEFAULT_QUERY_PROMPT, queryCount, parserKey,...rest } = fields;
-		const outputParser = new LineListOutputParser();
+		const { retriever, llm, prompt = DEFAULT_QUERY_PROMPT, queryCount, parserKey, ...rest } = fields;
 		const llmChain = new LLMChain({
 			llm,
 			prompt,
-			outputParser
+			outputParser: new LineListOutputParser()
 		});
 		return new this({
 			retriever,
@@ -101,11 +98,10 @@ var MultiQueryRetriever = class extends BaseRetriever {
 		});
 	}
 	async _generateQueries(question, runManager) {
-		const response = await this.llmChain.call({
+		const lines = (await this.llmChain.call({
 			question,
 			queryCount: this.queryCount
-		}, runManager?.getChild());
-		const lines = response.text[this.parserKey] || [];
+		}, runManager?.getChild())).text[this.parserKey] || [];
 		if (this.verbose) console.log(`Generated queries: ${lines}`);
 		return lines;
 	}
@@ -123,8 +119,7 @@ var MultiQueryRetriever = class extends BaseRetriever {
 			const key = `${doc.pageContent}:${JSON.stringify(Object.entries(doc.metadata).sort())}`;
 			uniqueDocumentsDict[key] = doc;
 		}
-		const uniqueDocuments = Object.values(uniqueDocumentsDict);
-		return uniqueDocuments;
+		return Object.values(uniqueDocumentsDict);
 	}
 	async _getRelevantDocuments(question, runManager) {
 		const queries = await this._generateQueries(question, runManager);

package/dist/retrievers/multi_query.js.map

@@ -1 +1 @@
-{"version":3,"file":"multi_query.js","names":["text: string","fields: MultiQueryRetrieverInput","fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n      llm: BaseLanguageModelInterface;\n      prompt?: BasePromptTemplate;\n    }","question: string","runManager?: CallbackManagerForRetrieverRun","queries: string[]","documents: Document[]","uniqueDocumentsDict: { [key: string]: Document }"],"sources":["../../src/retrievers/multi_query.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport {\n  BaseRetriever,\n  type BaseRetrieverInput,\n  type BaseRetrieverInterface,\n} from \"@langchain/core/retrievers\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseOutputParser } from \"@langchain/core/output_parsers\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../chains/llm_chain.js\";\nimport type { BaseDocumentCompressor } from \"./document_compressors/index.js\";\n\ninterface LineList {\n  lines: string[];\n}\n\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\nexport type MultiDocs = Document<Record<string, any>>[];\n\nclass LineListOutputParser extends BaseOutputParser<LineList> {\n  static lc_name() {\n    return \"LineListOutputParser\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  async parse(text: string): Promise<LineList> {\n    const startKeyIndex = text.indexOf(\"<questions>\");\n    const endKeyIndex = text.indexOf(\"</questions>\");\n    const questionsStartIndex =\n      startKeyIndex === -1 ? 0 : startKeyIndex + \"<questions>\".length;\n    const questionsEndIndex = endKeyIndex === -1 ? text.length : endKeyIndex;\n    const lines = text\n      .slice(questionsStartIndex, questionsEndIndex)\n      .trim()\n      .split(\"\\n\")\n      .filter((line) => line.trim() !== \"\");\n    return { lines };\n  }\n\n  getFormatInstructions(): string {\n    throw new Error(\"Not implemented.\");\n  }\n}\n\n// Create template\nconst DEFAULT_QUERY_PROMPT = /* #__PURE__ */ new PromptTemplate({\n  inputVariables: [\"question\", \"queryCount\"],\n  template: `You are an AI language model assistant. Your task is\nto generate {queryCount} different versions of the given user\nquestion to retrieve relevant documents from a vector database.\nBy generating multiple perspectives on the user question,\nyour goal is to help the user overcome some of the limitations\nof distance-based similarity search.\n\nProvide these alternative questions separated by newlines between XML tags. For example:\n\n<questions>\nQuestion 1\nQuestion 2\nQuestion 3\n</questions>\n\nOriginal question: {question}`,\n});\n\nexport interface MultiQueryRetrieverInput extends BaseRetrieverInput {\n  retriever: BaseRetrieverInterface;\n  /** @deprecated Pass a custom prompt into `.fromLLM` instead. */\n  llmChain: LLMChain<LineList>;\n  queryCount?: number;\n  parserKey?: string;\n  documentCompressor?: BaseDocumentCompressor | undefined;\n  documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;\n}\n\n/**\n * @example\n * ```typescript\n * const retriever = new MultiQueryRetriever.fromLLM({\n *   llm: new ChatAnthropic({}),\n *   retriever: new MemoryVectorStore().asRetriever(),\n *   verbose: true,\n * });\n * const retrievedDocs = await retriever.invoke(\n *   \"What are mitochondria made of?\",\n * );\n * ```\n */\nexport class MultiQueryRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"MultiQueryRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  private retriever: BaseRetrieverInterface;\n\n  private llmChain: LLMChain<LineList>;\n\n  private queryCount = 3;\n\n  private parserKey = \"lines\";\n\n  documentCompressor: BaseDocumentCompressor | undefined;\n\n  documentCompressorFilteringFn?: MultiQueryRetrieverInput[\"documentCompressorFilteringFn\"];\n\n  constructor(fields: MultiQueryRetrieverInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llmChain = fields.llmChain;\n    this.queryCount = fields.queryCount ?? this.queryCount;\n    this.parserKey = fields.parserKey ?? this.parserKey;\n    this.documentCompressor = fields.documentCompressor;\n    this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;\n  }\n\n  static fromLLM(\n    fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n      llm: BaseLanguageModelInterface;\n      prompt?: BasePromptTemplate;\n    }\n  ): MultiQueryRetriever {\n    const {\n      retriever,\n      llm,\n      prompt = DEFAULT_QUERY_PROMPT,\n      queryCount,\n      parserKey,\n      ...rest\n    } = fields;\n    const outputParser = new LineListOutputParser();\n    const llmChain = new LLMChain({ llm, prompt, outputParser });\n    return new this({ retriever, llmChain, queryCount, parserKey, ...rest });\n  }\n\n  // Generate the different queries for each retrieval, using our llmChain\n  private async _generateQueries(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<string[]> {\n    const response = await this.llmChain.call(\n      { question, queryCount: this.queryCount },\n      runManager?.getChild()\n    );\n    const lines = response.text[this.parserKey] || [];\n    if (this.verbose) {\n      console.log(`Generated queries: ${lines}`);\n    }\n    return lines;\n  }\n\n  // Retrieve documents using the original retriever\n  private async _retrieveDocuments(\n    queries: string[],\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const documents: Document[] = [];\n    await Promise.all(\n      queries.map(async (query) => {\n        const docs = await this.retriever.invoke(query, runManager?.getChild());\n        documents.push(...docs);\n      })\n    );\n    return documents;\n  }\n\n  // Deduplicate the documents that were returned in multiple retrievals\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const uniqueDocumentsDict: { [key: string]: Document } = {};\n\n    for (const doc of documents) {\n      const key = `${doc.pageContent}:${JSON.stringify(\n        Object.entries(doc.metadata).sort()\n      )}`;\n      uniqueDocumentsDict[key] = doc;\n    }\n\n    const uniqueDocuments = Object.values(uniqueDocumentsDict);\n    return uniqueDocuments;\n  }\n\n  async _getRelevantDocuments(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const queries = await this._generateQueries(question, runManager);\n    const documents = await this._retrieveDocuments(queries, runManager);\n    const uniqueDocuments = this._uniqueUnion(documents);\n\n    let outputDocs = uniqueDocuments;\n    if (this.documentCompressor && uniqueDocuments.length) {\n      outputDocs = await this.documentCompressor.compressDocuments(\n        uniqueDocuments,\n        question,\n        runManager?.getChild()\n      );\n      if (this.documentCompressorFilteringFn) {\n        outputDocs = this.documentCompressorFilteringFn(outputDocs);\n      }\n    }\n\n    return outputDocs;\n  }\n}\n"],"mappings":";;;;;;;;;AAoBA,IAAM,uBAAN,cAAmC,iBAA2B;CAC5D,OAAO,UAAU;AACf,SAAO;CACR;CAED,eAAe;EAAC;EAAa;EAAc;CAAa;CAExD,MAAM,MAAMA,MAAiC;EAC3C,MAAM,gBAAgB,KAAK,QAAQ,cAAc;EACjD,MAAM,cAAc,KAAK,QAAQ,eAAe;EAChD,MAAM,sBACJ,kBAAkB,KAAK,IAAI,gBAAgB;EAC7C,MAAM,oBAAoB,gBAAgB,KAAK,KAAK,SAAS;EAC7D,MAAM,QAAQ,KACX,MAAM,qBAAqB,kBAAkB,CAC7C,MAAM,CACN,MAAM,KAAK,CACX,OAAO,CAAC,SAAS,KAAK,MAAM,KAAK,GAAG;AACvC,SAAO,EAAE,MAAO;CACjB;CAED,wBAAgC;AAC9B,QAAM,IAAI,MAAM;CACjB;AACF;AAGD,MAAM,uCAAuC,IAAI,eAAe;CAC9D,gBAAgB,CAAC,YAAY,YAAa;CAC1C,UAAU,CAAC;;;;;;;;;;;;;;;6BAegB,CAAC;AAC7B;;;;;;;;;;;;;;AAyBD,IAAa,sBAAb,cAAyC,cAAc;CACrD,OAAO,UAAU;AACf,SAAO;CACR;CAED,eAAe;EAAC;EAAa;EAAc;CAAa;CAExD,AAAQ;CAER,AAAQ;CAER,AAAQ,aAAa;CAErB,AAAQ,YAAY;CAEpB;CAEA;CAEA,YAAYC,QAAkC;EAC5C,MAAM,OAAO;EACb,KAAK,YAAY,OAAO;EACxB,KAAK,WAAW,OAAO;EACvB,KAAK,aAAa,OAAO,cAAc,KAAK;EAC5C,KAAK,YAAY,OAAO,aAAa,KAAK;EAC1C,KAAK,qBAAqB,OAAO;EACjC,KAAK,gCAAgC,OAAO;CAC7C;CAED,OAAO,QACLC,QAIqB;EACrB,MAAM,EACJ,WACA,KACA,SAAS,sBACT,YACA,UACA,GAAG,MACJ,GAAG;EACJ,MAAM,eAAe,IAAI;EACzB,MAAM,WAAW,IAAI,SAAS;GAAE;GAAK;GAAQ;EAAc;AAC3D,SAAO,IAAI,KAAK;GAAE;GAAW;GAAU;GAAY;GAAW,GAAG;EAAM;CACxE;CAGD,MAAc,iBACZC,UACAC,YACmB;EACnB,MAAM,WAAW,MAAM,KAAK,SAAS,KACnC;GAAE;GAAU,YAAY,KAAK;EAAY,GACzC,YAAY,UAAU,CACvB;EACD,MAAM,QAAQ,SAAS,KAAK,KAAK,cAAc,CAAE;AACjD,MAAI,KAAK,SACP,QAAQ,IAAI,CAAC,mBAAmB,EAAE,OAAO,CAAC;AAE5C,SAAO;CACR;CAGD,MAAc,mBACZC,SACAD,YACqB;EACrB,MAAME,YAAwB,CAAE;EAChC,MAAM,QAAQ,IACZ,QAAQ,IAAI,OAAO,UAAU;GAC3B,MAAM,OAAO,MAAM,KAAK,UAAU,OAAO,OAAO,YAAY,UAAU,CAAC;GACvE,UAAU,KAAK,GAAG,KAAK;EACxB,EAAC,CACH;AACD,SAAO;CACR;CAGD,AAAQ,aAAaA,WAAmC;EACtD,MAAMC,sBAAmD,CAAE;AAE3D,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,EAAE,KAAK,UACrC,OAAO,QAAQ,IAAI,SAAS,CAAC,MAAM,CACpC,EAAE;GACH,oBAAoB,OAAO;EAC5B;EAED,MAAM,kBAAkB,OAAO,OAAO,oBAAoB;AAC1D,SAAO;CACR;CAED,MAAM,sBACJJ,UACAC,YACqB;EACrB,MAAM,UAAU,MAAM,KAAK,iBAAiB,UAAU,WAAW;EACjE,MAAM,YAAY,MAAM,KAAK,mBAAmB,SAAS,WAAW;EACpE,MAAM,kBAAkB,KAAK,aAAa,UAAU;EAEpD,IAAI,aAAa;AACjB,MAAI,KAAK,sBAAsB,gBAAgB,QAAQ;GACrD,aAAa,MAAM,KAAK,mBAAmB,kBACzC,iBACA,UACA,YAAY,UAAU,CACvB;AACD,OAAI,KAAK,+BACP,aAAa,KAAK,8BAA8B,WAAW;EAE9D;AAED,SAAO;CACR;AACF"}
+{"version":3,"file":"multi_query.js","names":[],"sources":["../../src/retrievers/multi_query.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport {\n  BaseRetriever,\n  type BaseRetrieverInput,\n  type BaseRetrieverInterface,\n} from \"@langchain/core/retrievers\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseOutputParser } from \"@langchain/core/output_parsers\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\nimport { LLMChain } from \"../chains/llm_chain.js\";\nimport type { BaseDocumentCompressor } from \"./document_compressors/index.js\";\n\ninterface LineList {\n  lines: string[];\n}\n\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\nexport type MultiDocs = Document<Record<string, any>>[];\n\nclass LineListOutputParser extends BaseOutputParser<LineList> {\n  static lc_name() {\n    return \"LineListOutputParser\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  async parse(text: string): Promise<LineList> {\n    const startKeyIndex = text.indexOf(\"<questions>\");\n    const endKeyIndex = text.indexOf(\"</questions>\");\n    const questionsStartIndex =\n      startKeyIndex === -1 ? 0 : startKeyIndex + \"<questions>\".length;\n    const questionsEndIndex = endKeyIndex === -1 ? text.length : endKeyIndex;\n    const lines = text\n      .slice(questionsStartIndex, questionsEndIndex)\n      .trim()\n      .split(\"\\n\")\n      .filter((line) => line.trim() !== \"\");\n    return { lines };\n  }\n\n  getFormatInstructions(): string {\n    throw new Error(\"Not implemented.\");\n  }\n}\n\n// Create template\nconst DEFAULT_QUERY_PROMPT = /* #__PURE__ */ new PromptTemplate({\n  inputVariables: [\"question\", \"queryCount\"],\n  template: `You are an AI language model assistant. Your task is\nto generate {queryCount} different versions of the given user\nquestion to retrieve relevant documents from a vector database.\nBy generating multiple perspectives on the user question,\nyour goal is to help the user overcome some of the limitations\nof distance-based similarity search.\n\nProvide these alternative questions separated by newlines between XML tags. For example:\n\n<questions>\nQuestion 1\nQuestion 2\nQuestion 3\n</questions>\n\nOriginal question: {question}`,\n});\n\nexport interface MultiQueryRetrieverInput extends BaseRetrieverInput {\n  retriever: BaseRetrieverInterface;\n  /** @deprecated Pass a custom prompt into `.fromLLM` instead. */\n  llmChain: LLMChain<LineList>;\n  queryCount?: number;\n  parserKey?: string;\n  documentCompressor?: BaseDocumentCompressor | undefined;\n  documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;\n}\n\n/**\n * @example\n * ```typescript\n * const retriever = new MultiQueryRetriever.fromLLM({\n *   llm: new ChatAnthropic({}),\n *   retriever: new MemoryVectorStore().asRetriever(),\n *   verbose: true,\n * });\n * const retrievedDocs = await retriever.invoke(\n *   \"What are mitochondria made of?\",\n * );\n * ```\n */\nexport class MultiQueryRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"MultiQueryRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multiquery\"];\n\n  private retriever: BaseRetrieverInterface;\n\n  private llmChain: LLMChain<LineList>;\n\n  private queryCount = 3;\n\n  private parserKey = \"lines\";\n\n  documentCompressor: BaseDocumentCompressor | undefined;\n\n  documentCompressorFilteringFn?: MultiQueryRetrieverInput[\"documentCompressorFilteringFn\"];\n\n  constructor(fields: MultiQueryRetrieverInput) {\n    super(fields);\n    this.retriever = fields.retriever;\n    this.llmChain = fields.llmChain;\n    this.queryCount = fields.queryCount ?? this.queryCount;\n    this.parserKey = fields.parserKey ?? this.parserKey;\n    this.documentCompressor = fields.documentCompressor;\n    this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;\n  }\n\n  static fromLLM(\n    fields: Omit<MultiQueryRetrieverInput, \"llmChain\"> & {\n      llm: BaseLanguageModelInterface;\n      prompt?: BasePromptTemplate;\n    }\n  ): MultiQueryRetriever {\n    const {\n      retriever,\n      llm,\n      prompt = DEFAULT_QUERY_PROMPT,\n      queryCount,\n      parserKey,\n      ...rest\n    } = fields;\n    const outputParser = new LineListOutputParser();\n    const llmChain = new LLMChain({ llm, prompt, outputParser });\n    return new this({ retriever, llmChain, queryCount, parserKey, ...rest });\n  }\n\n  // Generate the different queries for each retrieval, using our llmChain\n  private async _generateQueries(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<string[]> {\n    const response = await this.llmChain.call(\n      { question, queryCount: this.queryCount },\n      runManager?.getChild()\n    );\n    const lines = response.text[this.parserKey] || [];\n    if (this.verbose) {\n      console.log(`Generated queries: ${lines}`);\n    }\n    return lines;\n  }\n\n  // Retrieve documents using the original retriever\n  private async _retrieveDocuments(\n    queries: string[],\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const documents: Document[] = [];\n    await Promise.all(\n      queries.map(async (query) => {\n        const docs = await this.retriever.invoke(query, runManager?.getChild());\n        documents.push(...docs);\n      })\n    );\n    return documents;\n  }\n\n  // Deduplicate the documents that were returned in multiple retrievals\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const uniqueDocumentsDict: { [key: string]: Document } = {};\n\n    for (const doc of documents) {\n      const key = `${doc.pageContent}:${JSON.stringify(\n        Object.entries(doc.metadata).sort()\n      )}`;\n      uniqueDocumentsDict[key] = doc;\n    }\n\n    const uniqueDocuments = Object.values(uniqueDocumentsDict);\n    return uniqueDocuments;\n  }\n\n  async _getRelevantDocuments(\n    question: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    const queries = await this._generateQueries(question, runManager);\n    const documents = await this._retrieveDocuments(queries, runManager);\n    const uniqueDocuments = this._uniqueUnion(documents);\n\n    let outputDocs = uniqueDocuments;\n    if (this.documentCompressor && uniqueDocuments.length) {\n      outputDocs = await this.documentCompressor.compressDocuments(\n        uniqueDocuments,\n        question,\n        runManager?.getChild()\n      );\n      if (this.documentCompressorFilteringFn) {\n        outputDocs = this.documentCompressorFilteringFn(outputDocs);\n      }\n    }\n\n    return outputDocs;\n  }\n}\n"],"mappings":";;;;;;;;AAoBA,IAAM,uBAAN,cAAmC,iBAA2B;CAC5D,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAa;CAExD,MAAM,MAAM,MAAiC;EAC3C,MAAM,gBAAgB,KAAK,QAAQ,cAAc;EACjD,MAAM,cAAc,KAAK,QAAQ,eAAe;EAChD,MAAM,sBACJ,kBAAkB,KAAK,IAAI,gBAAgB;EAC7C,MAAM,oBAAoB,gBAAgB,KAAK,KAAK,SAAS;AAM7D,SAAO,EAAE,OALK,KACX,MAAM,qBAAqB,kBAAkB,CAC7C,MAAM,CACN,MAAM,KAAK,CACX,QAAQ,SAAS,KAAK,MAAM,KAAK,GAAG,EACvB;;CAGlB,wBAAgC;AAC9B,QAAM,IAAI,MAAM,mBAAmB;;;AAKvC,MAAM,uCAAuC,IAAI,eAAe;CAC9D,gBAAgB,CAAC,YAAY,aAAa;CAC1C,UAAU;;;;;;;;;;;;;;;;CAgBX,CAAC;;;;;;;;;;;;;;AAyBF,IAAa,sBAAb,cAAyC,cAAc;CACrD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAa;CAExD,AAAQ;CAER,AAAQ;CAER,AAAQ,aAAa;CAErB,AAAQ,YAAY;CAEpB;CAEA;CAEA,YAAY,QAAkC;AAC5C,QAAM,OAAO;AACb,OAAK,YAAY,OAAO;AACxB,OAAK,WAAW,OAAO;AACvB,OAAK,aAAa,OAAO,cAAc,KAAK;AAC5C,OAAK,YAAY,OAAO,aAAa,KAAK;AAC1C,OAAK,qBAAqB,OAAO;AACjC,OAAK,gCAAgC,OAAO;;CAG9C,OAAO,QACL,QAIqB;EACrB,MAAM,EACJ,WACA,KACA,SAAS,sBACT,YACA,WACA,GAAG,SACD;EAEJ,MAAM,WAAW,IAAI,SAAS;GAAE;GAAK;GAAQ,cADxB,IAAI,sBAAsB;GACY,CAAC;AAC5D,SAAO,IAAI,KAAK;GAAE;GAAW;GAAU;GAAY;GAAW,GAAG;GAAM,CAAC;;CAI1E,MAAc,iBACZ,UACA,YACmB;EAKnB,MAAM,SAJW,MAAM,KAAK,SAAS,KACnC;GAAE;GAAU,YAAY,KAAK;GAAY,EACzC,YAAY,UAAU,CACvB,EACsB,KAAK,KAAK,cAAc,EAAE;AACjD,MAAI,KAAK,QACP,SAAQ,IAAI,sBAAsB,QAAQ;AAE5C,SAAO;;CAIT,MAAc,mBACZ,SACA,YACqB;EACrB,MAAM,YAAwB,EAAE;AAChC,QAAM,QAAQ,IACZ,QAAQ,IAAI,OAAO,UAAU;GAC3B,MAAM,OAAO,MAAM,KAAK,UAAU,OAAO,OAAO,YAAY,UAAU,CAAC;AACvE,aAAU,KAAK,GAAG,KAAK;IACvB,CACH;AACD,SAAO;;CAIT,AAAQ,aAAa,WAAmC;EACtD,MAAM,sBAAmD,EAAE;AAE3D,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,GAAG,IAAI,YAAY,GAAG,KAAK,UACrC,OAAO,QAAQ,IAAI,SAAS,CAAC,MAAM,CACpC;AACD,uBAAoB,OAAO;;AAI7B,SADwB,OAAO,OAAO,oBAAoB;;CAI5D,MAAM,sBACJ,UACA,YACqB;EACrB,MAAM,UAAU,MAAM,KAAK,iBAAiB,UAAU,WAAW;EACjE,MAAM,YAAY,MAAM,KAAK,mBAAmB,SAAS,WAAW;EACpE,MAAM,kBAAkB,KAAK,aAAa,UAAU;EAEpD,IAAI,aAAa;AACjB,MAAI,KAAK,sBAAsB,gBAAgB,QAAQ;AACrD,gBAAa,MAAM,KAAK,mBAAmB,kBACzC,iBACA,UACA,YAAY,UAAU,CACvB;AACD,OAAI,KAAK,8BACP,cAAa,KAAK,8BAA8B,WAAW;;AAI/D,SAAO"}

package/dist/retrievers/multi_vector.cjs

@@ -1,10 +1,10 @@
-const require_rolldown_runtime = require('../_virtual/rolldown_runtime.cjs');
+Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });
+const require_runtime = require('../_virtual/_rolldown/runtime.cjs');
 const require_storage_encoder_backed = require('../storage/encoder_backed.cjs');
-const __langchain_core_retrievers = require_rolldown_runtime.__toESM(require("@langchain/core/retrievers"));
+let _langchain_core_retrievers = require("@langchain/core/retrievers");
 
 //#region src/retrievers/multi_vector.ts
-var multi_vector_exports = {};
-require_rolldown_runtime.__export(multi_vector_exports, { MultiVectorRetriever: () => MultiVectorRetriever });
+var multi_vector_exports = /* @__PURE__ */ require_runtime.__exportAll({ MultiVectorRetriever: () => MultiVectorRetriever });
 /**
 * A retriever that retrieves documents from a vector store and a document
 * store. It uses the vector store to find relevant documents based on a
@@ -23,7 +23,7 @@ require_rolldown_runtime.__export(multi_vector_exports, { MultiVectorRetriever:
 * console.log(retrieverResult[0].pageContent.length);
 * ```
 */
-var MultiVectorRetriever = class extends __langchain_core_retrievers.BaseRetriever {
+var MultiVectorRetriever = class extends _langchain_core_retrievers.BaseRetriever {
 	static lc_name() {
 		return "MultiVectorRetriever";
 	}
@@ -51,8 +51,7 @@ var MultiVectorRetriever = class extends __langchain_core_retrievers.BaseRetriev
 		const subDocs = await this.vectorstore.similaritySearch(query, this.childK);
 		const ids = [];
 		for (const doc of subDocs) if (doc.metadata[this.idKey] && !ids.includes(doc.metadata[this.idKey])) ids.push(doc.metadata[this.idKey]);
-		const docs = await this.docstore.mget(ids);
-		return docs.filter((doc) => doc !== void 0).slice(0, this.parentK);
+		return (await this.docstore.mget(ids)).filter((doc) => doc !== void 0).slice(0, this.parentK);
 	}
 };
 

package/dist/retrievers/multi_vector.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"multi_vector.cjs","names":["BaseRetriever","args: MultiVectorRetrieverInput","createDocumentStoreFromByteStore","query: string","ids: string[]"],"sources":["../../src/retrievers/multi_vector.ts"],"sourcesContent":["import {\n  BaseRetriever,\n  type BaseRetrieverInput,\n} from \"@langchain/core/retrievers\";\nimport type { VectorStoreInterface } from \"@langchain/core/vectorstores\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseStore, type BaseStoreInterface } from \"@langchain/core/stores\";\nimport { createDocumentStoreFromByteStore } from \"../storage/encoder_backed.js\";\n\n/**\n * Arguments for the MultiVectorRetriever class.\n */\nexport interface MultiVectorRetrieverInput extends BaseRetrieverInput {\n  vectorstore: VectorStoreInterface;\n  /** @deprecated Prefer `byteStore`. */\n  docstore?: BaseStoreInterface<string, Document>;\n  byteStore?: BaseStore<string, Uint8Array>;\n  idKey?: string;\n  childK?: number;\n  parentK?: number;\n}\n\n/**\n * A retriever that retrieves documents from a vector store and a document\n * store. It uses the vector store to find relevant documents based on a\n * query, and then retrieves the full documents from the document store.\n * @example\n * ```typescript\n * const retriever = new MultiVectorRetriever({\n *   vectorstore: new FaissStore(),\n *   byteStore: new InMemoryStore<Unit8Array>(),\n *   idKey: \"doc_id\",\n *   childK: 20,\n *   parentK: 5,\n * });\n *\n * const retrieverResult = await retriever.invoke(\"justice breyer\");\n * console.log(retrieverResult[0].pageContent.length);\n * ```\n */\nexport class MultiVectorRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"MultiVectorRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multi_vector\"];\n\n  public vectorstore: VectorStoreInterface;\n\n  public docstore: BaseStoreInterface<string, Document>;\n\n  protected idKey: string;\n\n  protected childK?: number;\n\n  protected parentK?: number;\n\n  constructor(args: MultiVectorRetrieverInput) {\n    super(args);\n    this.vectorstore = args.vectorstore;\n    if (args.byteStore) {\n      this.docstore = createDocumentStoreFromByteStore(args.byteStore);\n    } else if (args.docstore) {\n      this.docstore = args.docstore;\n    } else {\n      throw new Error(\n        \"byteStore and docstore are undefined. Please provide at least one.\"\n      );\n    }\n    this.idKey = args.idKey ?? \"doc_id\";\n    this.childK = args.childK;\n    this.parentK = args.parentK;\n  }\n\n  async _getRelevantDocuments(query: string): Promise<Document[]> {\n    const subDocs = await this.vectorstore.similaritySearch(query, this.childK);\n    const ids: string[] = [];\n    for (const doc of subDocs) {\n      if (doc.metadata[this.idKey] && !ids.includes(doc.metadata[this.idKey])) {\n        ids.push(doc.metadata[this.idKey]);\n      }\n    }\n    const docs = await this.docstore.mget(ids);\n    return docs\n      .filter((doc) => doc !== undefined)\n      .slice(0, this.parentK) as Document[];\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AAwCA,IAAa,uBAAb,cAA0CA,0CAAc;CACtD,OAAO,UAAU;AACf,SAAO;CACR;CAED,eAAe;EAAC;EAAa;EAAc;CAAe;CAE1D,AAAO;CAEP,AAAO;CAEP,AAAU;CAEV,AAAU;CAEV,AAAU;CAEV,YAAYC,MAAiC;EAC3C,MAAM,KAAK;EACX,KAAK,cAAc,KAAK;AACxB,MAAI,KAAK,WACP,KAAK,WAAWC,gEAAiC,KAAK,UAAU;WACvD,KAAK,UACd,KAAK,WAAW,KAAK;MAErB,OAAM,IAAI,MACR;EAGJ,KAAK,QAAQ,KAAK,SAAS;EAC3B,KAAK,SAAS,KAAK;EACnB,KAAK,UAAU,KAAK;CACrB;CAED,MAAM,sBAAsBC,OAAoC;EAC9D,MAAM,UAAU,MAAM,KAAK,YAAY,iBAAiB,OAAO,KAAK,OAAO;EAC3E,MAAMC,MAAgB,CAAE;AACxB,OAAK,MAAM,OAAO,QAChB,KAAI,IAAI,SAAS,KAAK,UAAU,CAAC,IAAI,SAAS,IAAI,SAAS,KAAK,OAAO,EACrE,IAAI,KAAK,IAAI,SAAS,KAAK,OAAO;EAGtC,MAAM,OAAO,MAAM,KAAK,SAAS,KAAK,IAAI;AAC1C,SAAO,KACJ,OAAO,CAAC,QAAQ,QAAQ,OAAU,CAClC,MAAM,GAAG,KAAK,QAAQ;CAC1B;AACF"}
+{"version":3,"file":"multi_vector.cjs","names":["BaseRetriever","createDocumentStoreFromByteStore"],"sources":["../../src/retrievers/multi_vector.ts"],"sourcesContent":["import {\n  BaseRetriever,\n  type BaseRetrieverInput,\n} from \"@langchain/core/retrievers\";\nimport type { VectorStoreInterface } from \"@langchain/core/vectorstores\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseStore, type BaseStoreInterface } from \"@langchain/core/stores\";\nimport { createDocumentStoreFromByteStore } from \"../storage/encoder_backed.js\";\n\n/**\n * Arguments for the MultiVectorRetriever class.\n */\nexport interface MultiVectorRetrieverInput extends BaseRetrieverInput {\n  vectorstore: VectorStoreInterface;\n  /** @deprecated Prefer `byteStore`. */\n  docstore?: BaseStoreInterface<string, Document>;\n  byteStore?: BaseStore<string, Uint8Array>;\n  idKey?: string;\n  childK?: number;\n  parentK?: number;\n}\n\n/**\n * A retriever that retrieves documents from a vector store and a document\n * store. It uses the vector store to find relevant documents based on a\n * query, and then retrieves the full documents from the document store.\n * @example\n * ```typescript\n * const retriever = new MultiVectorRetriever({\n *   vectorstore: new FaissStore(),\n *   byteStore: new InMemoryStore<Unit8Array>(),\n *   idKey: \"doc_id\",\n *   childK: 20,\n *   parentK: 5,\n * });\n *\n * const retrieverResult = await retriever.invoke(\"justice breyer\");\n * console.log(retrieverResult[0].pageContent.length);\n * ```\n */\nexport class MultiVectorRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"MultiVectorRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"multi_vector\"];\n\n  public vectorstore: VectorStoreInterface;\n\n  public docstore: BaseStoreInterface<string, Document>;\n\n  protected idKey: string;\n\n  protected childK?: number;\n\n  protected parentK?: number;\n\n  constructor(args: MultiVectorRetrieverInput) {\n    super(args);\n    this.vectorstore = args.vectorstore;\n    if (args.byteStore) {\n      this.docstore = createDocumentStoreFromByteStore(args.byteStore);\n    } else if (args.docstore) {\n      this.docstore = args.docstore;\n    } else {\n      throw new Error(\n        \"byteStore and docstore are undefined. Please provide at least one.\"\n      );\n    }\n    this.idKey = args.idKey ?? \"doc_id\";\n    this.childK = args.childK;\n    this.parentK = args.parentK;\n  }\n\n  async _getRelevantDocuments(query: string): Promise<Document[]> {\n    const subDocs = await this.vectorstore.similaritySearch(query, this.childK);\n    const ids: string[] = [];\n    for (const doc of subDocs) {\n      if (doc.metadata[this.idKey] && !ids.includes(doc.metadata[this.idKey])) {\n        ids.push(doc.metadata[this.idKey]);\n      }\n    }\n    const docs = await this.docstore.mget(ids);\n    return docs\n      .filter((doc) => doc !== undefined)\n      .slice(0, this.parentK) as Document[];\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AAwCA,IAAa,uBAAb,cAA0CA,yCAAc;CACtD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAe;CAE1D,AAAO;CAEP,AAAO;CAEP,AAAU;CAEV,AAAU;CAEV,AAAU;CAEV,YAAY,MAAiC;AAC3C,QAAM,KAAK;AACX,OAAK,cAAc,KAAK;AACxB,MAAI,KAAK,UACP,MAAK,WAAWC,gEAAiC,KAAK,UAAU;WACvD,KAAK,SACd,MAAK,WAAW,KAAK;MAErB,OAAM,IAAI,MACR,qEACD;AAEH,OAAK,QAAQ,KAAK,SAAS;AAC3B,OAAK,SAAS,KAAK;AACnB,OAAK,UAAU,KAAK;;CAGtB,MAAM,sBAAsB,OAAoC;EAC9D,MAAM,UAAU,MAAM,KAAK,YAAY,iBAAiB,OAAO,KAAK,OAAO;EAC3E,MAAM,MAAgB,EAAE;AACxB,OAAK,MAAM,OAAO,QAChB,KAAI,IAAI,SAAS,KAAK,UAAU,CAAC,IAAI,SAAS,IAAI,SAAS,KAAK,OAAO,CACrE,KAAI,KAAK,IAAI,SAAS,KAAK,OAAO;AAItC,UADa,MAAM,KAAK,SAAS,KAAK,IAAI,EAEvC,QAAQ,QAAQ,QAAQ,OAAU,CAClC,MAAM,GAAG,KAAK,QAAQ"}

package/dist/retrievers/multi_vector.d.cts

@@ -4,7 +4,6 @@ import { Document } from "@langchain/core/documents";
 import { BaseStore, BaseStoreInterface } from "@langchain/core/stores";
 
 //#region src/retrievers/multi_vector.d.ts
-
 /**
  * Arguments for the MultiVectorRetriever class.
  */

package/dist/retrievers/multi_vector.d.cts.map

@@ -1 +1 @@
-{"version":3,"file":"multi_vector.d.cts","names":["BaseRetriever","BaseRetrieverInput","VectorStoreInterface","Document","BaseStore","BaseStoreInterface","MultiVectorRetrieverInput","Uint8Array","MultiVectorRetriever","Promise"],"sources":["../../src/retrievers/multi_vector.d.ts"],"sourcesContent":["import { BaseRetriever, type BaseRetrieverInput } from \"@langchain/core/retrievers\";\nimport type { VectorStoreInterface } from \"@langchain/core/vectorstores\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseStore, type BaseStoreInterface } from \"@langchain/core/stores\";\n/**\n * Arguments for the MultiVectorRetriever class.\n */\nexport interface MultiVectorRetrieverInput extends BaseRetrieverInput {\n    vectorstore: VectorStoreInterface;\n    /** @deprecated Prefer `byteStore`. */\n    docstore?: BaseStoreInterface<string, Document>;\n    byteStore?: BaseStore<string, Uint8Array>;\n    idKey?: string;\n    childK?: number;\n    parentK?: number;\n}\n/**\n * A retriever that retrieves documents from a vector store and a document\n * store. It uses the vector store to find relevant documents based on a\n * query, and then retrieves the full documents from the document store.\n * @example\n * ```typescript\n * const retriever = new MultiVectorRetriever({\n *   vectorstore: new FaissStore(),\n *   byteStore: new InMemoryStore<Unit8Array>(),\n *   idKey: \"doc_id\",\n *   childK: 20,\n *   parentK: 5,\n * });\n *\n * const retrieverResult = await retriever.invoke(\"justice breyer\");\n * console.log(retrieverResult[0].pageContent.length);\n * ```\n */\nexport declare class MultiVectorRetriever extends BaseRetriever {\n    static lc_name(): string;\n    lc_namespace: string[];\n    vectorstore: VectorStoreInterface;\n    docstore: BaseStoreInterface<string, Document>;\n    protected idKey: string;\n    protected childK?: number;\n    protected parentK?: number;\n    constructor(args: MultiVectorRetrieverInput);\n    _getRelevantDocuments(query: string): Promise<Document[]>;\n}\n//# sourceMappingURL=multi_vector.d.ts.map"],"mappings":";;;;;;;;;AAOA;AACiBE,UADAI,yBAAAA,SAAkCL,kBAClCC,CAAAA;EAEyBC,WAAAA,EAFzBD,oBAEyBC;EAA3BE;EACmBE,QAAAA,CAAAA,EADnBF,kBACmBE,CAAAA,MAAAA,EADQJ,QACRI,CAAAA;EAAlBH,SAAAA,CAAAA,EAAAA,SAAAA,CAAAA,MAAAA,EAAkBG,UAAlBH,CAAAA;EAJmCH,KAAAA,CAAAA,EAAAA,MAAAA;EAAkB,MAAA,CAAA,EAAA,MAAA;EA2BhDO,OAAAA,CAAAA,EAAAA,MAAAA;;;;;;;;AAA0C;;;;;;;;;;;;cAA1CA,oBAAAA,SAA6BR,aAAAA;;;eAGjCE;YACHG,2BAA2BF;;;;oBAInBG;wCACoBG,QAAQN"}
+{"version":3,"file":"multi_vector.d.cts","names":[],"sources":["../../src/retrievers/multi_vector.ts"],"mappings":";;;;;;;;AAYA;UAAiB,yBAAA,SAAkC,kBAAA;EACjD,WAAA,EAAa,oBAAA;EAAA;EAEb,QAAA,GAAW,kBAAA,SAA2B,QAAA;EACtC,SAAA,GAAY,SAAA,SAAkB,UAAA;EAC9B,KAAA;EACA,MAAA;EACA,OAAA;AAAA;;;;;;;;;;;;;;;;AAqBF;;;cAAa,oBAAA,SAA6B,aAAA;EAAA,OACjC,OAAA,CAAA;EAIP,YAAA;EAEO,WAAA,EAAa,oBAAA;EAEb,QAAA,EAAU,kBAAA,SAA2B,QAAA;EAAA,UAElC,KAAA;EAAA,UAEA,MAAA;EAAA,UAEA,OAAA;EAEV,WAAA,CAAY,IAAA,EAAM,yBAAA;EAiBZ,qBAAA,CAAsB,KAAA,WAAgB,OAAA,CAAQ,QAAA;AAAA"}

package/dist/retrievers/multi_vector.d.ts

@@ -4,7 +4,6 @@ import { BaseRetriever, BaseRetrieverInput } from "@langchain/core/retrievers";
 import { VectorStoreInterface } from "@langchain/core/vectorstores";
 
 //#region src/retrievers/multi_vector.d.ts
-
 /**
  * Arguments for the MultiVectorRetriever class.
  */

package/dist/retrievers/multi_vector.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"multi_vector.d.ts","names":["BaseRetriever","BaseRetrieverInput","VectorStoreInterface","Document","BaseStore","BaseStoreInterface","MultiVectorRetrieverInput","Uint8Array","MultiVectorRetriever","Promise"],"sources":["../../src/retrievers/multi_vector.d.ts"],"sourcesContent":["import { BaseRetriever, type BaseRetrieverInput } from \"@langchain/core/retrievers\";\nimport type { VectorStoreInterface } from \"@langchain/core/vectorstores\";\nimport { Document } from \"@langchain/core/documents\";\nimport { BaseStore, type BaseStoreInterface } from \"@langchain/core/stores\";\n/**\n * Arguments for the MultiVectorRetriever class.\n */\nexport interface MultiVectorRetrieverInput extends BaseRetrieverInput {\n    vectorstore: VectorStoreInterface;\n    /** @deprecated Prefer `byteStore`. */\n    docstore?: BaseStoreInterface<string, Document>;\n    byteStore?: BaseStore<string, Uint8Array>;\n    idKey?: string;\n    childK?: number;\n    parentK?: number;\n}\n/**\n * A retriever that retrieves documents from a vector store and a document\n * store. It uses the vector store to find relevant documents based on a\n * query, and then retrieves the full documents from the document store.\n * @example\n * ```typescript\n * const retriever = new MultiVectorRetriever({\n *   vectorstore: new FaissStore(),\n *   byteStore: new InMemoryStore<Unit8Array>(),\n *   idKey: \"doc_id\",\n *   childK: 20,\n *   parentK: 5,\n * });\n *\n * const retrieverResult = await retriever.invoke(\"justice breyer\");\n * console.log(retrieverResult[0].pageContent.length);\n * ```\n */\nexport declare class MultiVectorRetriever extends BaseRetriever {\n    static lc_name(): string;\n    lc_namespace: string[];\n    vectorstore: VectorStoreInterface;\n    docstore: BaseStoreInterface<string, Document>;\n    protected idKey: string;\n    protected childK?: number;\n    protected parentK?: number;\n    constructor(args: MultiVectorRetrieverInput);\n    _getRelevantDocuments(query: string): Promise<Document[]>;\n}\n//# sourceMappingURL=multi_vector.d.ts.map"],"mappings":";;;;;;;;;AAOA;AACiBE,UADAI,yBAAAA,SAAkCL,kBAClCC,CAAAA;EAEyBC,WAAAA,EAFzBD,oBAEyBC;EAA3BE;EACmBE,QAAAA,CAAAA,EADnBF,kBACmBE,CAAAA,MAAAA,EADQJ,QACRI,CAAAA;EAAlBH,SAAAA,CAAAA,EAAAA,SAAAA,CAAAA,MAAAA,EAAkBG,UAAlBH,CAAAA;EAJmCH,KAAAA,CAAAA,EAAAA,MAAAA;EAAkB,MAAA,CAAA,EAAA,MAAA;EA2BhDO,OAAAA,CAAAA,EAAAA,MAAAA;;;;;;;;AAA0C;;;;;;;;;;;;cAA1CA,oBAAAA,SAA6BR,aAAAA;;;eAGjCE;YACHG,2BAA2BF;;;;oBAInBG;wCACoBG,QAAQN"}
+{"version":3,"file":"multi_vector.d.ts","names":[],"sources":["../../src/retrievers/multi_vector.ts"],"mappings":";;;;;;;;AAYA;UAAiB,yBAAA,SAAkC,kBAAA;EACjD,WAAA,EAAa,oBAAA;EAAA;EAEb,QAAA,GAAW,kBAAA,SAA2B,QAAA;EACtC,SAAA,GAAY,SAAA,SAAkB,UAAA;EAC9B,KAAA;EACA,MAAA;EACA,OAAA;AAAA;;;;;;;;;;;;;;;;AAqBF;;;cAAa,oBAAA,SAA6B,aAAA;EAAA,OACjC,OAAA,CAAA;EAIP,YAAA;EAEO,WAAA,EAAa,oBAAA;EAEb,QAAA,EAAU,kBAAA,SAA2B,QAAA;EAAA,UAElC,KAAA;EAAA,UAEA,MAAA;EAAA,UAEA,OAAA;EAEV,WAAA,CAAY,IAAA,EAAM,yBAAA;EAiBZ,qBAAA,CAAsB,KAAA,WAAgB,OAAA,CAAQ,QAAA;AAAA"}