@langchain/classic 1.0.0-alpha.1

package/dist/retrievers/matryoshka_retriever.js

@@ -0,0 +1,110 @@
+import { __export } from "../_virtual/rolldown_runtime.js";
+import { VectorStoreRetriever } from "@langchain/core/vectorstores";
+import { cosineSimilarity, euclideanDistance, innerProduct } from "@langchain/core/utils/math";
+
+//#region src/retrievers/matryoshka_retriever.ts
+var matryoshka_retriever_exports = {};
+__export(matryoshka_retriever_exports, { 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"
+* blog post {@link https://supabase.com/blog/matryoshka-embeddings}.
+*
+*
+* This class performs "Adaptive Retrieval" for searching text embeddings efficiently using the
+* Matryoshka Representation Learning (MRL) technique. It retrieves documents similar to a query
+* embedding in two steps:
+*
+* First-pass: Uses a lower dimensional sub-vector from the MRL embedding for an initial, fast,
+* but less accurate search.
+*
+* Second-pass: Re-ranks the top results from the first pass using the full, high-dimensional
+* embedding for higher accuracy.
+*
+*
+* This code implements MRL embeddings for efficient vector search by combining faster,
+* lower-dimensional initial search with accurate, high-dimensional re-ranking.
+*/
+var MatryoshkaRetriever = class extends VectorStoreRetriever {
+	smallK = 50;
+	largeK = 8;
+	largeEmbeddingKey = "lc_large_embedding";
+	largeEmbeddingModel;
+	searchType = "cosine";
+	constructor(fields) {
+		super(fields);
+		this.smallK = fields.smallK ?? this.smallK;
+		this.largeK = fields.largeK ?? this.largeK;
+		this.largeEmbeddingKey = fields.largeEmbeddingKey ?? this.largeEmbeddingKey;
+		this.largeEmbeddingModel = fields.largeEmbeddingModel;
+		this.searchType = fields.searchType ?? this.searchType;
+	}
+	/**
+	* Ranks documents based on their similarity to a query embedding using larger embeddings.
+	*
+	* This method takes a query embedding and a list of documents (smallResults) as input. Each document
+	* in the smallResults array has previously been associated with a large embedding stored in its metadata.
+	* Depending on the `searchType` (cosine, innerProduct, or euclidean), it calculates the similarity scores
+	* between the query embedding and each document's large embedding. It then ranks the documents based on
+	* these similarity scores, from the most similar to the least similar.
+	*
+	* The method returns a promise that resolves to an array of the top `largeK` documents, where `largeK`
+	* is a class property defining the number of documents to return. This subset of documents is determined
+	* by sorting the entire list of documents based on their similarity scores and then selecting the top
+	* `largeK` documents.
+	*
+	* @param {number[]} embeddedQuery The embedding of the query, represented as an array of numbers.
+	* @param {DocumentInterface[]} smallResults An array of documents, each with metadata that includes a large embedding for similarity comparison.
+	* @returns {Promise<DocumentInterface[]>} A promise that resolves to an array of the top `largeK` ranked documents based on their similarity to the query embedding.
+	*/
+	_rankByLargeEmbeddings(embeddedQuery, smallResults) {
+		const largeEmbeddings = smallResults.map((doc) => JSON.parse(doc.metadata[this.largeEmbeddingKey]));
+		let func;
+		switch (this.searchType) {
+			case "cosine":
+				func = () => cosineSimilarity([embeddedQuery], largeEmbeddings);
+				break;
+			case "innerProduct":
+				func = () => innerProduct([embeddedQuery], largeEmbeddings);
+				break;
+			case "euclidean":
+				func = () => euclideanDistance([embeddedQuery], largeEmbeddings);
+				break;
+			default: throw new Error(`Unknown search type: ${this.searchType}`);
+		}
+		const [similarityScores] = func();
+		let indices = Array.from({ length: smallResults.length }, (_, index) => index);
+		indices = indices.map((v, i) => [similarityScores[i], v]).sort(([a], [b]) => b - a).slice(0, this.largeK).map(([, i]) => i);
+		return indices.map((i) => smallResults[i]);
+	}
+	async _getRelevantDocuments(query) {
+		const [embeddedQuery, smallResults] = await Promise.all([this.largeEmbeddingModel.embedQuery(query), this.vectorStore.similaritySearch(query, this.smallK, this.filter)]);
+		return this._rankByLargeEmbeddings(embeddedQuery, smallResults);
+	}
+	/**
+	* Override the default `addDocuments` method to embed the documents twice,
+	* once using the larger embeddings model, and then again using the default
+	* embedding model linked to the vector store.
+	*
+	* @param {DocumentInterface[]} documents - An array of documents to add to the vector store.
+	* @param {AddDocumentOptions} options - An optional object containing additional options for adding documents.
+	* @returns {Promise<string[] | void>} A promise that resolves to an array of the document IDs that were added to the vector store.
+	*/
+	addDocuments = async (documents, options) => {
+		if (documents.some((doc) => this.largeEmbeddingKey in doc.metadata)) throw new Error(`All documents must not contain the large embedding key: ${this.largeEmbeddingKey} in their metadata.`);
+		const allDocPageContent = documents.map((doc) => doc.pageContent);
+		const allDocLargeEmbeddings = await this.largeEmbeddingModel.embedDocuments(allDocPageContent);
+		const newDocuments = documents.map((doc, idx) => ({
+			...doc,
+			metadata: {
+				...doc.metadata,
+				[this.largeEmbeddingKey]: JSON.stringify(allDocLargeEmbeddings[idx])
+			}
+		}));
+		return this.vectorStore.addDocuments(newDocuments, options);
+	};
+};
+
+//#endregion
+export { MatryoshkaRetriever, matryoshka_retriever_exports };
+//# sourceMappingURL=matryoshka_retriever.js.map

package/dist/retrievers/matryoshka_retriever.js.map

@@ -0,0 +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 = await this.largeEmbeddingModel.embedDocuments(\n      allDocPageContent\n    );\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,wBAAwB,MAAM,KAAK,oBAAoB,eAC3D,kBACD;EAED,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"}

package/dist/retrievers/multi_query.cjs

@@ -0,0 +1,150 @@
+const require_rolldown_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"));
+
+//#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 {
+	static lc_name() {
+		return "LineListOutputParser";
+	}
+	lc_namespace = [
+		"langchain",
+		"retrievers",
+		"multiquery"
+	];
+	async parse(text) {
+		const startKeyIndex = text.indexOf("<questions>");
+		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 };
+	}
+	getFormatInstructions() {
+		throw new Error("Not implemented.");
+	}
+};
+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
+question to retrieve relevant documents from a vector database.
+By generating multiple perspectives on the user question,
+your goal is to help the user overcome some of the limitations
+of distance-based similarity search.
+
+Provide these alternative questions separated by newlines between XML tags. For example:
+
+<questions>
+Question 1
+Question 2
+Question 3
+</questions>
+
+Original question: {question}`
+});
+/**
+* @example
+* ```typescript
+* const retriever = new MultiQueryRetriever.fromLLM({
+*   llm: new ChatAnthropic({}),
+*   retriever: new MemoryVectorStore().asRetriever(),
+*   verbose: true,
+* });
+* const retrievedDocs = await retriever.invoke(
+*   "What are mitochondria made of?",
+* );
+* ```
+*/
+var MultiQueryRetriever = class extends __langchain_core_retrievers.BaseRetriever {
+	static lc_name() {
+		return "MultiQueryRetriever";
+	}
+	lc_namespace = [
+		"langchain",
+		"retrievers",
+		"multiquery"
+	];
+	retriever;
+	llmChain;
+	queryCount = 3;
+	parserKey = "lines";
+	documentCompressor;
+	documentCompressorFilteringFn;
+	constructor(fields) {
+		super(fields);
+		this.retriever = fields.retriever;
+		this.llmChain = fields.llmChain;
+		this.queryCount = fields.queryCount ?? this.queryCount;
+		this.parserKey = fields.parserKey ?? this.parserKey;
+		this.documentCompressor = fields.documentCompressor;
+		this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;
+	}
+	static fromLLM(fields) {
+		const { retriever, llm, prompt = DEFAULT_QUERY_PROMPT, queryCount, parserKey,...rest } = fields;
+		const outputParser = new LineListOutputParser();
+		const llmChain = new require_llm_chain.LLMChain({
+			llm,
+			prompt,
+			outputParser
+		});
+		return new this({
+			retriever,
+			llmChain,
+			queryCount,
+			parserKey,
+			...rest
+		});
+	}
+	async _generateQueries(question, runManager) {
+		const response = await this.llmChain.call({
+			question,
+			queryCount: this.queryCount
+		}, runManager?.getChild());
+		const lines = response.text[this.parserKey] || [];
+		if (this.verbose) console.log(`Generated queries: ${lines}`);
+		return lines;
+	}
+	async _retrieveDocuments(queries, runManager) {
+		const documents = [];
+		await Promise.all(queries.map(async (query) => {
+			const docs = await this.retriever.invoke(query, runManager?.getChild());
+			documents.push(...docs);
+		}));
+		return documents;
+	}
+	_uniqueUnion(documents) {
+		const uniqueDocumentsDict = {};
+		for (const doc of documents) {
+			const key = `${doc.pageContent}:${JSON.stringify(Object.entries(doc.metadata).sort())}`;
+			uniqueDocumentsDict[key] = doc;
+		}
+		const uniqueDocuments = Object.values(uniqueDocumentsDict);
+		return uniqueDocuments;
+	}
+	async _getRelevantDocuments(question, runManager) {
+		const queries = await this._generateQueries(question, runManager);
+		const documents = await this._retrieveDocuments(queries, runManager);
+		const uniqueDocuments = this._uniqueUnion(documents);
+		let outputDocs = uniqueDocuments;
+		if (this.documentCompressor && uniqueDocuments.length) {
+			outputDocs = await this.documentCompressor.compressDocuments(uniqueDocuments, question, runManager?.getChild());
+			if (this.documentCompressorFilteringFn) outputDocs = this.documentCompressorFilteringFn(outputDocs);
+		}
+		return outputDocs;
+	}
+};
+
+//#endregion
+exports.MultiQueryRetriever = MultiQueryRetriever;
+Object.defineProperty(exports, 'multi_query_exports', {
+  enumerable: true,
+  get: function () {
+    return multi_query_exports;
+  }
+});
+//# sourceMappingURL=multi_query.cjs.map

package/dist/retrievers/multi_query.cjs.map

@@ -0,0 +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"}

package/dist/retrievers/multi_query.d.cts

@@ -0,0 +1,59 @@
+import { LLMChain } from "../chains/llm_chain.cjs";
+import { BaseDocumentCompressor } from "./document_compressors/index.cjs";
+import { BasePromptTemplate } from "@langchain/core/prompts";
+import { BaseLanguageModelInterface } from "@langchain/core/language_models/base";
+import { CallbackManagerForRetrieverRun } from "@langchain/core/callbacks/manager";
+import { BaseRetriever, BaseRetrieverInput, BaseRetrieverInterface } from "@langchain/core/retrievers";
+import { Document } from "@langchain/core/documents";
+
+//#region src/retrievers/multi_query.d.ts
+interface LineList {
+  lines: string[];
+}
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+type MultiDocs = Document<Record<string, any>>[];
+interface MultiQueryRetrieverInput extends BaseRetrieverInput {
+  retriever: BaseRetrieverInterface;
+  /** @deprecated Pass a custom prompt into `.fromLLM` instead. */
+  llmChain: LLMChain<LineList>;
+  queryCount?: number;
+  parserKey?: string;
+  documentCompressor?: BaseDocumentCompressor | undefined;
+  documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;
+}
+/**
+ * @example
+ * ```typescript
+ * const retriever = new MultiQueryRetriever.fromLLM({
+ *   llm: new ChatAnthropic({}),
+ *   retriever: new MemoryVectorStore().asRetriever(),
+ *   verbose: true,
+ * });
+ * const retrievedDocs = await retriever.invoke(
+ *   "What are mitochondria made of?",
+ * );
+ * ```
+ */
+declare class MultiQueryRetriever extends BaseRetriever {
+  static lc_name(): string;
+  lc_namespace: string[];
+  private retriever;
+  private llmChain;
+  private queryCount;
+  private parserKey;
+  documentCompressor: BaseDocumentCompressor | undefined;
+  documentCompressorFilteringFn?: MultiQueryRetrieverInput["documentCompressorFilteringFn"];
+  constructor(fields: MultiQueryRetrieverInput);
+  static fromLLM(fields: Omit<MultiQueryRetrieverInput, "llmChain"> & {
+    llm: BaseLanguageModelInterface;
+    prompt?: BasePromptTemplate;
+  }): MultiQueryRetriever;
+  private _generateQueries;
+  private _retrieveDocuments;
+  // Deduplicate the documents that were returned in multiple retrievals
+  private _uniqueUnion;
+  _getRelevantDocuments(question: string, runManager?: CallbackManagerForRetrieverRun): Promise<Document[]>;
+}
+//#endregion
+export { MultiDocs, MultiQueryRetriever, MultiQueryRetrieverInput };
+//# sourceMappingURL=multi_query.d.cts.map

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

@@ -0,0 +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}\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\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    // Deduplicate the documents that were returned in multiple retrievals\n    private _uniqueUnion;\n    _getRelevantDocuments(question: string, runManager?: CallbackManagerForRetrieverRun): Promise<Document[]>;\n}\nexport {};\n"],"mappings":";;;;;;;;;UAOUS,QAAAA;;AADoE;AAK9E;AAAqB,KAATC,SAAAA,GAAYN,QAAH,CAAYO,MAAZ,CAAA,MAAA,EAAA,GAAA,CAAA,CAAA,EAAA;AAAYA,UAChBC,wBAAAA,SAAiCV,kBADjBS,CAAAA;EAAM,SAAfP,EAETD,sBAFSC;EAAQ;EACfQ,QAAAA,EAGHL,QAHGK,CAGMH,QAHkB,CAAA;EAAA,UAAA,CAAA,EAAA,MAAA;EAAA,SAC1BN,CAAAA,EAAAA,MAAAA;EAAsB,kBAEdM,CAAAA,EAGED,sBAHFC,GAAAA,SAAAA;EAAQ,6BAAjBF,CAAAA,EAAAA,CAAAA,IAAAA,EAI6BG,SAJ7BH,EAAAA,GAI2CG,SAJ3CH;;;;;AAHsD;AAsBpE;;;;;;;;;AAaQM,cAbaA,mBAAAA,SAA4BZ,aAAAA,CAazCY;EAAmB,OAK8BP,OAAAA,CAAAA,CAAAA,EAAAA,MAAAA;EAA8B,YAAWF,EAAAA,MAAAA,EAAAA;EAAQ,QAAhBW,SAAAA;EAAO,QAlBhDd,QAAAA;EAAa,QAAA,UAAA;;sBAOtCO;kCACYI;sBACZA;yBACGE,KAAKF;SACnBZ;aACIK;MACTQ;;;;;uDAKiDP,iCAAiCS,QAAQX"}

package/dist/retrievers/multi_query.d.ts

@@ -0,0 +1,59 @@
+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 { BaseRetriever, BaseRetrieverInput, BaseRetrieverInterface } from "@langchain/core/retrievers";
+
+//#region src/retrievers/multi_query.d.ts
+interface LineList {
+  lines: string[];
+}
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+type MultiDocs = Document<Record<string, any>>[];
+interface MultiQueryRetrieverInput extends BaseRetrieverInput {
+  retriever: BaseRetrieverInterface;
+  /** @deprecated Pass a custom prompt into `.fromLLM` instead. */
+  llmChain: LLMChain<LineList>;
+  queryCount?: number;
+  parserKey?: string;
+  documentCompressor?: BaseDocumentCompressor | undefined;
+  documentCompressorFilteringFn?: (docs: MultiDocs) => MultiDocs;
+}
+/**
+ * @example
+ * ```typescript
+ * const retriever = new MultiQueryRetriever.fromLLM({
+ *   llm: new ChatAnthropic({}),
+ *   retriever: new MemoryVectorStore().asRetriever(),
+ *   verbose: true,
+ * });
+ * const retrievedDocs = await retriever.invoke(
+ *   "What are mitochondria made of?",
+ * );
+ * ```
+ */
+declare class MultiQueryRetriever extends BaseRetriever {
+  static lc_name(): string;
+  lc_namespace: string[];
+  private retriever;
+  private llmChain;
+  private queryCount;
+  private parserKey;
+  documentCompressor: BaseDocumentCompressor | undefined;
+  documentCompressorFilteringFn?: MultiQueryRetrieverInput["documentCompressorFilteringFn"];
+  constructor(fields: MultiQueryRetrieverInput);
+  static fromLLM(fields: Omit<MultiQueryRetrieverInput, "llmChain"> & {
+    llm: BaseLanguageModelInterface;
+    prompt?: BasePromptTemplate;
+  }): MultiQueryRetriever;
+  private _generateQueries;
+  private _retrieveDocuments;
+  // Deduplicate the documents that were returned in multiple retrievals
+  private _uniqueUnion;
+  _getRelevantDocuments(question: string, runManager?: CallbackManagerForRetrieverRun): Promise<Document[]>;
+}
+//#endregion
+export { MultiDocs, MultiQueryRetriever, MultiQueryRetrieverInput };
+//# sourceMappingURL=multi_query.d.ts.map

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

@@ -0,0 +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}\n// eslint-disable-next-line @typescript-eslint/no-explicit-any\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    // Deduplicate the documents that were returned in multiple retrievals\n    private _uniqueUnion;\n    _getRelevantDocuments(question: string, runManager?: CallbackManagerForRetrieverRun): Promise<Document[]>;\n}\nexport {};\n"],"mappings":";;;;;;;;;UAOUS,QAAAA;;AADoE;AAK9E;AAAqB,KAATC,SAAAA,GAAYN,QAAH,CAAYO,MAAZ,CAAA,MAAA,EAAA,GAAA,CAAA,CAAA,EAAA;AAAYA,UAChBC,wBAAAA,SAAiCV,kBADjBS,CAAAA;EAAM,SAAfP,EAETD,sBAFSC;EAAQ;EACfQ,QAAAA,EAGHL,QAHGK,CAGMH,QAHkB,CAAA;EAAA,UAAA,CAAA,EAAA,MAAA;EAAA,SAC1BN,CAAAA,EAAAA,MAAAA;EAAsB,kBAEdM,CAAAA,EAGED,sBAHFC,GAAAA,SAAAA;EAAQ,6BAAjBF,CAAAA,EAAAA,CAAAA,IAAAA,EAI6BG,SAJ7BH,EAAAA,GAI2CG,SAJ3CH;;;;;AAHsD;AAsBpE;;;;;;;;;AAaQM,cAbaA,mBAAAA,SAA4BZ,aAAAA,CAazCY;EAAmB,OAK8BP,OAAAA,CAAAA,CAAAA,EAAAA,MAAAA;EAA8B,YAAWF,EAAAA,MAAAA,EAAAA;EAAQ,QAAhBW,SAAAA;EAAO,QAlBhDd,QAAAA;EAAa,QAAA,UAAA;;sBAOtCO;kCACYI;sBACZA;yBACGE,KAAKF;SACnBZ;aACIK;MACTQ;;;;;uDAKiDP,iCAAiCS,QAAQX"}

package/dist/retrievers/multi_query.js

@@ -0,0 +1,144 @@
+import { __export } 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 LineListOutputParser = class extends BaseOutputParser {
+	static lc_name() {
+		return "LineListOutputParser";
+	}
+	lc_namespace = [
+		"langchain",
+		"retrievers",
+		"multiquery"
+	];
+	async parse(text) {
+		const startKeyIndex = text.indexOf("<questions>");
+		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 };
+	}
+	getFormatInstructions() {
+		throw new Error("Not implemented.");
+	}
+};
+const DEFAULT_QUERY_PROMPT = /* @__PURE__ */ new PromptTemplate({
+	inputVariables: ["question", "queryCount"],
+	template: `You are an AI language model assistant. Your task is
+to generate {queryCount} different versions of the given user
+question to retrieve relevant documents from a vector database.
+By generating multiple perspectives on the user question,
+your goal is to help the user overcome some of the limitations
+of distance-based similarity search.
+
+Provide these alternative questions separated by newlines between XML tags. For example:
+
+<questions>
+Question 1
+Question 2
+Question 3
+</questions>
+
+Original question: {question}`
+});
+/**
+* @example
+* ```typescript
+* const retriever = new MultiQueryRetriever.fromLLM({
+*   llm: new ChatAnthropic({}),
+*   retriever: new MemoryVectorStore().asRetriever(),
+*   verbose: true,
+* });
+* const retrievedDocs = await retriever.invoke(
+*   "What are mitochondria made of?",
+* );
+* ```
+*/
+var MultiQueryRetriever = class extends BaseRetriever {
+	static lc_name() {
+		return "MultiQueryRetriever";
+	}
+	lc_namespace = [
+		"langchain",
+		"retrievers",
+		"multiquery"
+	];
+	retriever;
+	llmChain;
+	queryCount = 3;
+	parserKey = "lines";
+	documentCompressor;
+	documentCompressorFilteringFn;
+	constructor(fields) {
+		super(fields);
+		this.retriever = fields.retriever;
+		this.llmChain = fields.llmChain;
+		this.queryCount = fields.queryCount ?? this.queryCount;
+		this.parserKey = fields.parserKey ?? this.parserKey;
+		this.documentCompressor = fields.documentCompressor;
+		this.documentCompressorFilteringFn = fields.documentCompressorFilteringFn;
+	}
+	static fromLLM(fields) {
+		const { retriever, llm, prompt = DEFAULT_QUERY_PROMPT, queryCount, parserKey,...rest } = fields;
+		const outputParser = new LineListOutputParser();
+		const llmChain = new LLMChain({
+			llm,
+			prompt,
+			outputParser
+		});
+		return new this({
+			retriever,
+			llmChain,
+			queryCount,
+			parserKey,
+			...rest
+		});
+	}
+	async _generateQueries(question, runManager) {
+		const response = await this.llmChain.call({
+			question,
+			queryCount: this.queryCount
+		}, runManager?.getChild());
+		const lines = response.text[this.parserKey] || [];
+		if (this.verbose) console.log(`Generated queries: ${lines}`);
+		return lines;
+	}
+	async _retrieveDocuments(queries, runManager) {
+		const documents = [];
+		await Promise.all(queries.map(async (query) => {
+			const docs = await this.retriever.invoke(query, runManager?.getChild());
+			documents.push(...docs);
+		}));
+		return documents;
+	}
+	_uniqueUnion(documents) {
+		const uniqueDocumentsDict = {};
+		for (const doc of documents) {
+			const key = `${doc.pageContent}:${JSON.stringify(Object.entries(doc.metadata).sort())}`;
+			uniqueDocumentsDict[key] = doc;
+		}
+		const uniqueDocuments = Object.values(uniqueDocumentsDict);
+		return uniqueDocuments;
+	}
+	async _getRelevantDocuments(question, runManager) {
+		const queries = await this._generateQueries(question, runManager);
+		const documents = await this._retrieveDocuments(queries, runManager);
+		const uniqueDocuments = this._uniqueUnion(documents);
+		let outputDocs = uniqueDocuments;
+		if (this.documentCompressor && uniqueDocuments.length) {
+			outputDocs = await this.documentCompressor.compressDocuments(uniqueDocuments, question, runManager?.getChild());
+			if (this.documentCompressorFilteringFn) outputDocs = this.documentCompressorFilteringFn(outputDocs);
+		}
+		return outputDocs;
+	}
+};
+
+//#endregion
+export { MultiQueryRetriever, multi_query_exports };
+//# sourceMappingURL=multi_query.js.map

package/dist/retrievers/multi_query.js.map

@@ -0,0 +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"}