@langchain/classic 1.0.23 → 1.0.24-dev-1773698445534

package/dist/retrievers/ensemble.cjs

@@ -1,7 +1,6 @@
-Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });
-const require_runtime = require('../_virtual/_rolldown/runtime.cjs');
+Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
+const require_runtime = require("../_virtual/_rolldown/runtime.cjs");
 let _langchain_core_retrievers = require("@langchain/core/retrievers");
-
 //#region src/retrievers/ensemble.ts
 var ensemble_exports = /* @__PURE__ */ require_runtime.__exportAll({ EnsembleRetriever: () => EnsembleRetriever });
 /**
@@ -62,13 +61,13 @@ var EnsembleRetriever = class extends _langchain_core_retrievers.BaseRetriever {
 		return result;
 	}
 };
-
 //#endregion
 exports.EnsembleRetriever = EnsembleRetriever;
-Object.defineProperty(exports, 'ensemble_exports', {
-  enumerable: true,
-  get: function () {
-    return ensemble_exports;
-  }
+Object.defineProperty(exports, "ensemble_exports", {
+	enumerable: true,
+	get: function() {
+		return ensemble_exports;
+	}
 });
+
 //# sourceMappingURL=ensemble.cjs.map

package/dist/retrievers/ensemble.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"ensemble.cjs","names":["BaseRetriever"],"sources":["../../src/retrievers/ensemble.ts"],"sourcesContent":["import { BaseRetriever, BaseRetrieverInput } from \"@langchain/core/retrievers\";\nimport { Document, DocumentInterface } from \"@langchain/core/documents\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\nexport interface EnsembleRetrieverInput extends BaseRetrieverInput {\n  /** A list of retrievers to ensemble. */\n  retrievers: BaseRetriever[];\n  /**\n   * A list of weights corresponding to the retrievers. Defaults to equal\n   * weighting for all retrievers.\n   */\n  weights?: number[];\n  /**\n   * A constant added to the rank, controlling the balance between the importance\n   * of high-ranked items and the consideration given to lower-ranked items.\n   * Default is 60.\n   */\n  c?: number;\n}\n\n/**\n * Ensemble retriever that aggregates and orders the results of\n * multiple retrievers by using weighted Reciprocal Rank Fusion.\n */\nexport class EnsembleRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"EnsembleRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"ensemble_retriever\"];\n\n  retrievers: BaseRetriever[];\n\n  weights: number[];\n\n  c = 60;\n\n  constructor(args: EnsembleRetrieverInput) {\n    super(args);\n    this.retrievers = args.retrievers;\n    this.weights =\n      args.weights ||\n      new Array(args.retrievers.length).fill(1 / args.retrievers.length);\n    this.c = args.c || 60;\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    return this._rankFusion(query, runManager);\n  }\n\n  async _rankFusion(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    const retrieverDocs = await Promise.all(\n      this.retrievers.map((retriever, i) =>\n        retriever.invoke(query, {\n          callbacks: runManager?.getChild(`retriever_${i + 1}`),\n        })\n      )\n    );\n\n    const fusedDocs = await this._weightedReciprocalRank(retrieverDocs);\n    return fusedDocs;\n  }\n\n  async _weightedReciprocalRank(docList: DocumentInterface[][]) {\n    if (docList.length !== this.weights.length) {\n      throw new Error(\n        \"Number of retrieved document lists must be equal to the number of weights.\"\n      );\n    }\n\n    const rrfScoreDict = docList.reduce(\n      (rffScore: Record<string, number>, retrieverDoc, idx) => {\n        let rank = 1;\n        const weight = this.weights[idx];\n        while (rank <= retrieverDoc.length) {\n          const { pageContent } = retrieverDoc[rank - 1];\n          if (!rffScore[pageContent]) {\n            rffScore[pageContent] = 0;\n          }\n          rffScore[pageContent] += weight / (rank + this.c);\n          rank += 1;\n        }\n\n        return rffScore;\n      },\n      {}\n    );\n\n    const uniqueDocs = this._uniqueUnion(docList.flat());\n    const sortedDocs = Array.from(uniqueDocs).sort(\n      (a, b) => rrfScoreDict[b.pageContent] - rrfScoreDict[a.pageContent]\n    );\n\n    return sortedDocs;\n  }\n\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const documentSet = new Set();\n    const result = [];\n\n    for (const doc of documents) {\n      const key = doc.pageContent;\n      if (!documentSet.has(key)) {\n        documentSet.add(key);\n        result.push(doc);\n      }\n    }\n\n    return result;\n  }\n}\n"],"mappings":";;;;;;;;;;AAwBA,IAAa,oBAAb,cAAuCA,yCAAc;CACnD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAqB;CAEhE;CAEA;CAEA,IAAI;CAEJ,YAAY,MAA8B;AACxC,QAAM,KAAK;AACX,OAAK,aAAa,KAAK;AACvB,OAAK,UACH,KAAK,WACL,IAAI,MAAM,KAAK,WAAW,OAAO,CAAC,KAAK,IAAI,KAAK,WAAW,OAAO;AACpE,OAAK,IAAI,KAAK,KAAK;;CAGrB,MAAM,sBACJ,OACA,YACA;AACA,SAAO,KAAK,YAAY,OAAO,WAAW;;CAG5C,MAAM,YACJ,OACA,YACA;EACA,MAAM,gBAAgB,MAAM,QAAQ,IAClC,KAAK,WAAW,KAAK,WAAW,MAC9B,UAAU,OAAO,OAAO,EACtB,WAAW,YAAY,SAAS,aAAa,IAAI,IAAI,EACtD,CAAC,CACH,CACF;AAGD,SADkB,MAAM,KAAK,wBAAwB,cAAc;;CAIrE,MAAM,wBAAwB,SAAgC;AAC5D,MAAI,QAAQ,WAAW,KAAK,QAAQ,OAClC,OAAM,IAAI,MACR,6EACD;EAGH,MAAM,eAAe,QAAQ,QAC1B,UAAkC,cAAc,QAAQ;GACvD,IAAI,OAAO;GACX,MAAM,SAAS,KAAK,QAAQ;AAC5B,UAAO,QAAQ,aAAa,QAAQ;IAClC,MAAM,EAAE,gBAAgB,aAAa,OAAO;AAC5C,QAAI,CAAC,SAAS,aACZ,UAAS,eAAe;AAE1B,aAAS,gBAAgB,UAAU,OAAO,KAAK;AAC/C,YAAQ;;AAGV,UAAO;KAET,EAAE,CACH;EAED,MAAM,aAAa,KAAK,aAAa,QAAQ,MAAM,CAAC;AAKpD,SAJmB,MAAM,KAAK,WAAW,CAAC,MACvC,GAAG,MAAM,aAAa,EAAE,eAAe,aAAa,EAAE,aACxD;;CAKH,AAAQ,aAAa,WAAmC;EACtD,MAAM,8BAAc,IAAI,KAAK;EAC7B,MAAM,SAAS,EAAE;AAEjB,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,IAAI;AAChB,OAAI,CAAC,YAAY,IAAI,IAAI,EAAE;AACzB,gBAAY,IAAI,IAAI;AACpB,WAAO,KAAK,IAAI;;;AAIpB,SAAO"}
+{"version":3,"file":"ensemble.cjs","names":["BaseRetriever"],"sources":["../../src/retrievers/ensemble.ts"],"sourcesContent":["import { BaseRetriever, BaseRetrieverInput } from \"@langchain/core/retrievers\";\nimport { Document, DocumentInterface } from \"@langchain/core/documents\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\nexport interface EnsembleRetrieverInput extends BaseRetrieverInput {\n  /** A list of retrievers to ensemble. */\n  retrievers: BaseRetriever[];\n  /**\n   * A list of weights corresponding to the retrievers. Defaults to equal\n   * weighting for all retrievers.\n   */\n  weights?: number[];\n  /**\n   * A constant added to the rank, controlling the balance between the importance\n   * of high-ranked items and the consideration given to lower-ranked items.\n   * Default is 60.\n   */\n  c?: number;\n}\n\n/**\n * Ensemble retriever that aggregates and orders the results of\n * multiple retrievers by using weighted Reciprocal Rank Fusion.\n */\nexport class EnsembleRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"EnsembleRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"ensemble_retriever\"];\n\n  retrievers: BaseRetriever[];\n\n  weights: number[];\n\n  c = 60;\n\n  constructor(args: EnsembleRetrieverInput) {\n    super(args);\n    this.retrievers = args.retrievers;\n    this.weights =\n      args.weights ||\n      new Array(args.retrievers.length).fill(1 / args.retrievers.length);\n    this.c = args.c || 60;\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    return this._rankFusion(query, runManager);\n  }\n\n  async _rankFusion(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    const retrieverDocs = await Promise.all(\n      this.retrievers.map((retriever, i) =>\n        retriever.invoke(query, {\n          callbacks: runManager?.getChild(`retriever_${i + 1}`),\n        })\n      )\n    );\n\n    const fusedDocs = await this._weightedReciprocalRank(retrieverDocs);\n    return fusedDocs;\n  }\n\n  async _weightedReciprocalRank(docList: DocumentInterface[][]) {\n    if (docList.length !== this.weights.length) {\n      throw new Error(\n        \"Number of retrieved document lists must be equal to the number of weights.\"\n      );\n    }\n\n    const rrfScoreDict = docList.reduce(\n      (rffScore: Record<string, number>, retrieverDoc, idx) => {\n        let rank = 1;\n        const weight = this.weights[idx];\n        while (rank <= retrieverDoc.length) {\n          const { pageContent } = retrieverDoc[rank - 1];\n          if (!rffScore[pageContent]) {\n            rffScore[pageContent] = 0;\n          }\n          rffScore[pageContent] += weight / (rank + this.c);\n          rank += 1;\n        }\n\n        return rffScore;\n      },\n      {}\n    );\n\n    const uniqueDocs = this._uniqueUnion(docList.flat());\n    const sortedDocs = Array.from(uniqueDocs).sort(\n      (a, b) => rrfScoreDict[b.pageContent] - rrfScoreDict[a.pageContent]\n    );\n\n    return sortedDocs;\n  }\n\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const documentSet = new Set();\n    const result = [];\n\n    for (const doc of documents) {\n      const key = doc.pageContent;\n      if (!documentSet.has(key)) {\n        documentSet.add(key);\n        result.push(doc);\n      }\n    }\n\n    return result;\n  }\n}\n"],"mappings":";;;;;;;;;AAwBA,IAAa,oBAAb,cAAuCA,2BAAAA,cAAc;CACnD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAqB;CAEhE;CAEA;CAEA,IAAI;CAEJ,YAAY,MAA8B;AACxC,QAAM,KAAK;AACX,OAAK,aAAa,KAAK;AACvB,OAAK,UACH,KAAK,WACL,IAAI,MAAM,KAAK,WAAW,OAAO,CAAC,KAAK,IAAI,KAAK,WAAW,OAAO;AACpE,OAAK,IAAI,KAAK,KAAK;;CAGrB,MAAM,sBACJ,OACA,YACA;AACA,SAAO,KAAK,YAAY,OAAO,WAAW;;CAG5C,MAAM,YACJ,OACA,YACA;EACA,MAAM,gBAAgB,MAAM,QAAQ,IAClC,KAAK,WAAW,KAAK,WAAW,MAC9B,UAAU,OAAO,OAAO,EACtB,WAAW,YAAY,SAAS,aAAa,IAAI,IAAI,EACtD,CAAC,CACH,CACF;AAGD,SADkB,MAAM,KAAK,wBAAwB,cAAc;;CAIrE,MAAM,wBAAwB,SAAgC;AAC5D,MAAI,QAAQ,WAAW,KAAK,QAAQ,OAClC,OAAM,IAAI,MACR,6EACD;EAGH,MAAM,eAAe,QAAQ,QAC1B,UAAkC,cAAc,QAAQ;GACvD,IAAI,OAAO;GACX,MAAM,SAAS,KAAK,QAAQ;AAC5B,UAAO,QAAQ,aAAa,QAAQ;IAClC,MAAM,EAAE,gBAAgB,aAAa,OAAO;AAC5C,QAAI,CAAC,SAAS,aACZ,UAAS,eAAe;AAE1B,aAAS,gBAAgB,UAAU,OAAO,KAAK;AAC/C,YAAQ;;AAGV,UAAO;KAET,EAAE,CACH;EAED,MAAM,aAAa,KAAK,aAAa,QAAQ,MAAM,CAAC;AAKpD,SAJmB,MAAM,KAAK,WAAW,CAAC,MACvC,GAAG,MAAM,aAAa,EAAE,eAAe,aAAa,EAAE,aACxD;;CAKH,aAAqB,WAAmC;EACtD,MAAM,8BAAc,IAAI,KAAK;EAC7B,MAAM,SAAS,EAAE;AAEjB,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,IAAI;AAChB,OAAI,CAAC,YAAY,IAAI,IAAI,EAAE;AACzB,gBAAY,IAAI,IAAI;AACpB,WAAO,KAAK,IAAI;;;AAIpB,SAAO"}

package/dist/retrievers/ensemble.js

@@ -1,6 +1,5 @@
 import { __exportAll } from "../_virtual/_rolldown/runtime.js";
 import { BaseRetriever } from "@langchain/core/retrievers";
-
 //#region src/retrievers/ensemble.ts
 var ensemble_exports = /* @__PURE__ */ __exportAll({ EnsembleRetriever: () => EnsembleRetriever });
 /**
@@ -61,7 +60,7 @@ var EnsembleRetriever = class extends BaseRetriever {
 		return result;
 	}
 };
-
 //#endregion
 export { EnsembleRetriever, ensemble_exports };
+
 //# sourceMappingURL=ensemble.js.map

package/dist/retrievers/ensemble.js.map

@@ -1 +1 @@
-{"version":3,"file":"ensemble.js","names":[],"sources":["../../src/retrievers/ensemble.ts"],"sourcesContent":["import { BaseRetriever, BaseRetrieverInput } from \"@langchain/core/retrievers\";\nimport { Document, DocumentInterface } from \"@langchain/core/documents\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\nexport interface EnsembleRetrieverInput extends BaseRetrieverInput {\n  /** A list of retrievers to ensemble. */\n  retrievers: BaseRetriever[];\n  /**\n   * A list of weights corresponding to the retrievers. Defaults to equal\n   * weighting for all retrievers.\n   */\n  weights?: number[];\n  /**\n   * A constant added to the rank, controlling the balance between the importance\n   * of high-ranked items and the consideration given to lower-ranked items.\n   * Default is 60.\n   */\n  c?: number;\n}\n\n/**\n * Ensemble retriever that aggregates and orders the results of\n * multiple retrievers by using weighted Reciprocal Rank Fusion.\n */\nexport class EnsembleRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"EnsembleRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"ensemble_retriever\"];\n\n  retrievers: BaseRetriever[];\n\n  weights: number[];\n\n  c = 60;\n\n  constructor(args: EnsembleRetrieverInput) {\n    super(args);\n    this.retrievers = args.retrievers;\n    this.weights =\n      args.weights ||\n      new Array(args.retrievers.length).fill(1 / args.retrievers.length);\n    this.c = args.c || 60;\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    return this._rankFusion(query, runManager);\n  }\n\n  async _rankFusion(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    const retrieverDocs = await Promise.all(\n      this.retrievers.map((retriever, i) =>\n        retriever.invoke(query, {\n          callbacks: runManager?.getChild(`retriever_${i + 1}`),\n        })\n      )\n    );\n\n    const fusedDocs = await this._weightedReciprocalRank(retrieverDocs);\n    return fusedDocs;\n  }\n\n  async _weightedReciprocalRank(docList: DocumentInterface[][]) {\n    if (docList.length !== this.weights.length) {\n      throw new Error(\n        \"Number of retrieved document lists must be equal to the number of weights.\"\n      );\n    }\n\n    const rrfScoreDict = docList.reduce(\n      (rffScore: Record<string, number>, retrieverDoc, idx) => {\n        let rank = 1;\n        const weight = this.weights[idx];\n        while (rank <= retrieverDoc.length) {\n          const { pageContent } = retrieverDoc[rank - 1];\n          if (!rffScore[pageContent]) {\n            rffScore[pageContent] = 0;\n          }\n          rffScore[pageContent] += weight / (rank + this.c);\n          rank += 1;\n        }\n\n        return rffScore;\n      },\n      {}\n    );\n\n    const uniqueDocs = this._uniqueUnion(docList.flat());\n    const sortedDocs = Array.from(uniqueDocs).sort(\n      (a, b) => rrfScoreDict[b.pageContent] - rrfScoreDict[a.pageContent]\n    );\n\n    return sortedDocs;\n  }\n\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const documentSet = new Set();\n    const result = [];\n\n    for (const doc of documents) {\n      const key = doc.pageContent;\n      if (!documentSet.has(key)) {\n        documentSet.add(key);\n        result.push(doc);\n      }\n    }\n\n    return result;\n  }\n}\n"],"mappings":";;;;;;;;;AAwBA,IAAa,oBAAb,cAAuC,cAAc;CACnD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAqB;CAEhE;CAEA;CAEA,IAAI;CAEJ,YAAY,MAA8B;AACxC,QAAM,KAAK;AACX,OAAK,aAAa,KAAK;AACvB,OAAK,UACH,KAAK,WACL,IAAI,MAAM,KAAK,WAAW,OAAO,CAAC,KAAK,IAAI,KAAK,WAAW,OAAO;AACpE,OAAK,IAAI,KAAK,KAAK;;CAGrB,MAAM,sBACJ,OACA,YACA;AACA,SAAO,KAAK,YAAY,OAAO,WAAW;;CAG5C,MAAM,YACJ,OACA,YACA;EACA,MAAM,gBAAgB,MAAM,QAAQ,IAClC,KAAK,WAAW,KAAK,WAAW,MAC9B,UAAU,OAAO,OAAO,EACtB,WAAW,YAAY,SAAS,aAAa,IAAI,IAAI,EACtD,CAAC,CACH,CACF;AAGD,SADkB,MAAM,KAAK,wBAAwB,cAAc;;CAIrE,MAAM,wBAAwB,SAAgC;AAC5D,MAAI,QAAQ,WAAW,KAAK,QAAQ,OAClC,OAAM,IAAI,MACR,6EACD;EAGH,MAAM,eAAe,QAAQ,QAC1B,UAAkC,cAAc,QAAQ;GACvD,IAAI,OAAO;GACX,MAAM,SAAS,KAAK,QAAQ;AAC5B,UAAO,QAAQ,aAAa,QAAQ;IAClC,MAAM,EAAE,gBAAgB,aAAa,OAAO;AAC5C,QAAI,CAAC,SAAS,aACZ,UAAS,eAAe;AAE1B,aAAS,gBAAgB,UAAU,OAAO,KAAK;AAC/C,YAAQ;;AAGV,UAAO;KAET,EAAE,CACH;EAED,MAAM,aAAa,KAAK,aAAa,QAAQ,MAAM,CAAC;AAKpD,SAJmB,MAAM,KAAK,WAAW,CAAC,MACvC,GAAG,MAAM,aAAa,EAAE,eAAe,aAAa,EAAE,aACxD;;CAKH,AAAQ,aAAa,WAAmC;EACtD,MAAM,8BAAc,IAAI,KAAK;EAC7B,MAAM,SAAS,EAAE;AAEjB,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,IAAI;AAChB,OAAI,CAAC,YAAY,IAAI,IAAI,EAAE;AACzB,gBAAY,IAAI,IAAI;AACpB,WAAO,KAAK,IAAI;;;AAIpB,SAAO"}
+{"version":3,"file":"ensemble.js","names":[],"sources":["../../src/retrievers/ensemble.ts"],"sourcesContent":["import { BaseRetriever, BaseRetrieverInput } from \"@langchain/core/retrievers\";\nimport { Document, DocumentInterface } from \"@langchain/core/documents\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\nexport interface EnsembleRetrieverInput extends BaseRetrieverInput {\n  /** A list of retrievers to ensemble. */\n  retrievers: BaseRetriever[];\n  /**\n   * A list of weights corresponding to the retrievers. Defaults to equal\n   * weighting for all retrievers.\n   */\n  weights?: number[];\n  /**\n   * A constant added to the rank, controlling the balance between the importance\n   * of high-ranked items and the consideration given to lower-ranked items.\n   * Default is 60.\n   */\n  c?: number;\n}\n\n/**\n * Ensemble retriever that aggregates and orders the results of\n * multiple retrievers by using weighted Reciprocal Rank Fusion.\n */\nexport class EnsembleRetriever extends BaseRetriever {\n  static lc_name() {\n    return \"EnsembleRetriever\";\n  }\n\n  lc_namespace = [\"langchain\", \"retrievers\", \"ensemble_retriever\"];\n\n  retrievers: BaseRetriever[];\n\n  weights: number[];\n\n  c = 60;\n\n  constructor(args: EnsembleRetrieverInput) {\n    super(args);\n    this.retrievers = args.retrievers;\n    this.weights =\n      args.weights ||\n      new Array(args.retrievers.length).fill(1 / args.retrievers.length);\n    this.c = args.c || 60;\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    return this._rankFusion(query, runManager);\n  }\n\n  async _rankFusion(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ) {\n    const retrieverDocs = await Promise.all(\n      this.retrievers.map((retriever, i) =>\n        retriever.invoke(query, {\n          callbacks: runManager?.getChild(`retriever_${i + 1}`),\n        })\n      )\n    );\n\n    const fusedDocs = await this._weightedReciprocalRank(retrieverDocs);\n    return fusedDocs;\n  }\n\n  async _weightedReciprocalRank(docList: DocumentInterface[][]) {\n    if (docList.length !== this.weights.length) {\n      throw new Error(\n        \"Number of retrieved document lists must be equal to the number of weights.\"\n      );\n    }\n\n    const rrfScoreDict = docList.reduce(\n      (rffScore: Record<string, number>, retrieverDoc, idx) => {\n        let rank = 1;\n        const weight = this.weights[idx];\n        while (rank <= retrieverDoc.length) {\n          const { pageContent } = retrieverDoc[rank - 1];\n          if (!rffScore[pageContent]) {\n            rffScore[pageContent] = 0;\n          }\n          rffScore[pageContent] += weight / (rank + this.c);\n          rank += 1;\n        }\n\n        return rffScore;\n      },\n      {}\n    );\n\n    const uniqueDocs = this._uniqueUnion(docList.flat());\n    const sortedDocs = Array.from(uniqueDocs).sort(\n      (a, b) => rrfScoreDict[b.pageContent] - rrfScoreDict[a.pageContent]\n    );\n\n    return sortedDocs;\n  }\n\n  private _uniqueUnion(documents: Document[]): Document[] {\n    const documentSet = new Set();\n    const result = [];\n\n    for (const doc of documents) {\n      const key = doc.pageContent;\n      if (!documentSet.has(key)) {\n        documentSet.add(key);\n        result.push(doc);\n      }\n    }\n\n    return result;\n  }\n}\n"],"mappings":";;;;;;;;AAwBA,IAAa,oBAAb,cAAuC,cAAc;CACnD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAqB;CAEhE;CAEA;CAEA,IAAI;CAEJ,YAAY,MAA8B;AACxC,QAAM,KAAK;AACX,OAAK,aAAa,KAAK;AACvB,OAAK,UACH,KAAK,WACL,IAAI,MAAM,KAAK,WAAW,OAAO,CAAC,KAAK,IAAI,KAAK,WAAW,OAAO;AACpE,OAAK,IAAI,KAAK,KAAK;;CAGrB,MAAM,sBACJ,OACA,YACA;AACA,SAAO,KAAK,YAAY,OAAO,WAAW;;CAG5C,MAAM,YACJ,OACA,YACA;EACA,MAAM,gBAAgB,MAAM,QAAQ,IAClC,KAAK,WAAW,KAAK,WAAW,MAC9B,UAAU,OAAO,OAAO,EACtB,WAAW,YAAY,SAAS,aAAa,IAAI,IAAI,EACtD,CAAC,CACH,CACF;AAGD,SADkB,MAAM,KAAK,wBAAwB,cAAc;;CAIrE,MAAM,wBAAwB,SAAgC;AAC5D,MAAI,QAAQ,WAAW,KAAK,QAAQ,OAClC,OAAM,IAAI,MACR,6EACD;EAGH,MAAM,eAAe,QAAQ,QAC1B,UAAkC,cAAc,QAAQ;GACvD,IAAI,OAAO;GACX,MAAM,SAAS,KAAK,QAAQ;AAC5B,UAAO,QAAQ,aAAa,QAAQ;IAClC,MAAM,EAAE,gBAAgB,aAAa,OAAO;AAC5C,QAAI,CAAC,SAAS,aACZ,UAAS,eAAe;AAE1B,aAAS,gBAAgB,UAAU,OAAO,KAAK;AAC/C,YAAQ;;AAGV,UAAO;KAET,EAAE,CACH;EAED,MAAM,aAAa,KAAK,aAAa,QAAQ,MAAM,CAAC;AAKpD,SAJmB,MAAM,KAAK,WAAW,CAAC,MACvC,GAAG,MAAM,aAAa,EAAE,eAAe,aAAa,EAAE,aACxD;;CAKH,aAAqB,WAAmC;EACtD,MAAM,8BAAc,IAAI,KAAK;EAC7B,MAAM,SAAS,EAAE;AAEjB,OAAK,MAAM,OAAO,WAAW;GAC3B,MAAM,MAAM,IAAI;AAChB,OAAI,CAAC,YAAY,IAAI,IAAI,EAAE;AACzB,gBAAY,IAAI,IAAI;AACpB,WAAO,KAAK,IAAI;;;AAIpB,SAAO"}

package/dist/retrievers/hyde.cjs

@@ -1,9 +1,8 @@
-Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });
-const require_runtime = require('../_virtual/_rolldown/runtime.cjs');
+Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
+const require_runtime = require("../_virtual/_rolldown/runtime.cjs");
 let _langchain_core_prompts = require("@langchain/core/prompts");
 let _langchain_core_prompt_values = require("@langchain/core/prompt_values");
 let _langchain_core_vectorstores = require("@langchain/core/vectorstores");
-
 //#region src/retrievers/hyde.ts
 var hyde_exports = /* @__PURE__ */ require_runtime.__exportAll({
 	HydeRetriever: () => HydeRetriever,
@@ -113,14 +112,14 @@ Passage:`;
 	}
 	return _langchain_core_prompts.PromptTemplate.fromTemplate(template);
 }
-
 //#endregion
 exports.HydeRetriever = HydeRetriever;
 exports.getPromptTemplateFromKey = getPromptTemplateFromKey;
-Object.defineProperty(exports, 'hyde_exports', {
-  enumerable: true,
-  get: function () {
-    return hyde_exports;
-  }
+Object.defineProperty(exports, "hyde_exports", {
+	enumerable: true,
+	get: function() {
+		return hyde_exports;
+	}
 });
+
 //# sourceMappingURL=hyde.cjs.map

package/dist/retrievers/hyde.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"hyde.cjs","names":["VectorStoreRetriever","StringPromptValue","PromptTemplate"],"sources":["../../src/retrievers/hyde.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { Document } from \"@langchain/core/documents\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport {\n  StringPromptValue,\n  BasePromptValue,\n} from \"@langchain/core/prompt_values\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\n/**\n * A string that corresponds to a specific prompt template.\n */\nexport type PromptKey =\n  | \"websearch\"\n  | \"scifact\"\n  | \"arguana\"\n  | \"trec-covid\"\n  | \"fiqa\"\n  | \"dbpedia-entity\"\n  | \"trec-news\"\n  | \"mr-tydi\";\n\n/**\n * Options for the HydeRetriever class, which includes a BaseLanguageModel\n * instance, a VectorStore instance, and an optional promptTemplate which\n * can either be a BasePromptTemplate instance or a PromptKey.\n */\nexport type HydeRetrieverOptions<V extends VectorStore> =\n  VectorStoreRetrieverInput<V> & {\n    llm: BaseLanguageModelInterface;\n    promptTemplate?: BasePromptTemplate | PromptKey;\n  };\n\n/**\n * A class for retrieving relevant documents based on a given query. It\n * extends the VectorStoreRetriever class and uses a BaseLanguageModel to\n * generate a hypothetical answer to the query, which is then used to\n * retrieve relevant documents.\n * @example\n * ```typescript\n * const retriever = new HydeRetriever({\n *   vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *   llm: new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *   k: 1,\n * });\n * await vectorStore.addDocuments(\n *   [\n *     \"My name is John.\",\n *     \"My name is Bob.\",\n *     \"My favourite food is pizza.\",\n *     \"My favourite food is pasta.\",\n *   ].map((pageContent) => new Document({ pageContent })),\n * );\n * const results = await retriever.invoke(\n *   \"What is my favourite food?\",\n * );\n * ```\n */\nexport class HydeRetriever<\n  V extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<V> {\n  static lc_name() {\n    return \"HydeRetriever\";\n  }\n\n  get lc_namespace(): string[] {\n    return [\"langchain\", \"retrievers\", \"hyde\"];\n  }\n\n  llm: BaseLanguageModelInterface;\n\n  promptTemplate?: BasePromptTemplate;\n\n  constructor(fields: HydeRetrieverOptions<V>) {\n    super(fields);\n    this.llm = fields.llm;\n    this.promptTemplate =\n      typeof fields.promptTemplate === \"string\"\n        ? getPromptTemplateFromKey(fields.promptTemplate)\n        : fields.promptTemplate;\n    if (this.promptTemplate) {\n      const { inputVariables } = this.promptTemplate;\n      if (inputVariables.length !== 1 && inputVariables[0] !== \"question\") {\n        throw new Error(\n          `Prompt template must accept a single input variable 'question'. Invalid input variables for prompt template: ${inputVariables}`\n        );\n      }\n    }\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    let value: BasePromptValue = new StringPromptValue(query);\n\n    // Use a custom template if provided\n    if (this.promptTemplate) {\n      value = await this.promptTemplate.formatPromptValue({ question: query });\n    }\n\n    // Get a hypothetical answer from the LLM\n    const res = await this.llm.generatePrompt([value]);\n    const answer = res.generations[0][0].text;\n\n    // Retrieve relevant documents based on the hypothetical answer\n    const results = await this.vectorStore.similaritySearch(\n      answer,\n      this.k,\n      this.filter,\n      runManager?.getChild(\"vectorstore\")\n    );\n\n    return results;\n  }\n}\n\n/**\n * Returns a BasePromptTemplate instance based on a given PromptKey.\n */\nexport function getPromptTemplateFromKey(key: PromptKey): BasePromptTemplate {\n  let template: string;\n\n  switch (key) {\n    case \"websearch\":\n      template = `Please write a passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"scifact\":\n      template = `Please write a scientific paper passage to support/refute the claim\nClaim: {question}\nPassage:`;\n      break;\n    case \"arguana\":\n      template = `Please write a counter argument for the passage\nPassage: {question}\nCounter Argument:`;\n      break;\n    case \"trec-covid\":\n      template = `Please write a scientific paper passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"fiqa\":\n      template = `Please write a financial article passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"dbpedia-entity\":\n      template = `Please write a passage to answer the question.\nQuestion: {question}\nPassage:`;\n      break;\n    case \"trec-news\":\n      template = `Please write a news passage about the topic.\nTopic: {question}\nPassage:`;\n      break;\n    case \"mr-tydi\":\n      template = `Please write a passage in Swahili/Korean/Japanese/Bengali to answer the question in detail.\nQuestion: {question}\nPassage:`;\n      break;\n    default:\n      throw new Error(`Invalid prompt key: ${key}`);\n  }\n\n  return PromptTemplate.fromTemplate(template);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+DA,IAAa,gBAAb,cAEUA,kDAAwB;CAChC,OAAO,UAAU;AACf,SAAO;;CAGT,IAAI,eAAyB;AAC3B,SAAO;GAAC;GAAa;GAAc;GAAO;;CAG5C;CAEA;CAEA,YAAY,QAAiC;AAC3C,QAAM,OAAO;AACb,OAAK,MAAM,OAAO;AAClB,OAAK,iBACH,OAAO,OAAO,mBAAmB,WAC7B,yBAAyB,OAAO,eAAe,GAC/C,OAAO;AACb,MAAI,KAAK,gBAAgB;GACvB,MAAM,EAAE,mBAAmB,KAAK;AAChC,OAAI,eAAe,WAAW,KAAK,eAAe,OAAO,WACvD,OAAM,IAAI,MACR,gHAAgH,iBACjH;;;CAKP,MAAM,sBACJ,OACA,YACqB;EACrB,IAAI,QAAyB,IAAIC,gDAAkB,MAAM;AAGzD,MAAI,KAAK,eACP,SAAQ,MAAM,KAAK,eAAe,kBAAkB,EAAE,UAAU,OAAO,CAAC;EAK1E,MAAM,UADM,MAAM,KAAK,IAAI,eAAe,CAAC,MAAM,CAAC,EAC/B,YAAY,GAAG,GAAG;AAUrC,SAPgB,MAAM,KAAK,YAAY,iBACrC,QACA,KAAK,GACL,KAAK,QACL,YAAY,SAAS,cAAc,CACpC;;;;;;AASL,SAAgB,yBAAyB,KAAoC;CAC3E,IAAI;AAEJ,SAAQ,KAAR;EACE,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,QACE,OAAM,IAAI,MAAM,uBAAuB,MAAM;;AAGjD,QAAOC,uCAAe,aAAa,SAAS"}
+{"version":3,"file":"hyde.cjs","names":["VectorStoreRetriever","StringPromptValue","PromptTemplate"],"sources":["../../src/retrievers/hyde.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { Document } from \"@langchain/core/documents\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport {\n  StringPromptValue,\n  BasePromptValue,\n} from \"@langchain/core/prompt_values\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\n/**\n * A string that corresponds to a specific prompt template.\n */\nexport type PromptKey =\n  | \"websearch\"\n  | \"scifact\"\n  | \"arguana\"\n  | \"trec-covid\"\n  | \"fiqa\"\n  | \"dbpedia-entity\"\n  | \"trec-news\"\n  | \"mr-tydi\";\n\n/**\n * Options for the HydeRetriever class, which includes a BaseLanguageModel\n * instance, a VectorStore instance, and an optional promptTemplate which\n * can either be a BasePromptTemplate instance or a PromptKey.\n */\nexport type HydeRetrieverOptions<V extends VectorStore> =\n  VectorStoreRetrieverInput<V> & {\n    llm: BaseLanguageModelInterface;\n    promptTemplate?: BasePromptTemplate | PromptKey;\n  };\n\n/**\n * A class for retrieving relevant documents based on a given query. It\n * extends the VectorStoreRetriever class and uses a BaseLanguageModel to\n * generate a hypothetical answer to the query, which is then used to\n * retrieve relevant documents.\n * @example\n * ```typescript\n * const retriever = new HydeRetriever({\n *   vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *   llm: new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *   k: 1,\n * });\n * await vectorStore.addDocuments(\n *   [\n *     \"My name is John.\",\n *     \"My name is Bob.\",\n *     \"My favourite food is pizza.\",\n *     \"My favourite food is pasta.\",\n *   ].map((pageContent) => new Document({ pageContent })),\n * );\n * const results = await retriever.invoke(\n *   \"What is my favourite food?\",\n * );\n * ```\n */\nexport class HydeRetriever<\n  V extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<V> {\n  static lc_name() {\n    return \"HydeRetriever\";\n  }\n\n  get lc_namespace(): string[] {\n    return [\"langchain\", \"retrievers\", \"hyde\"];\n  }\n\n  llm: BaseLanguageModelInterface;\n\n  promptTemplate?: BasePromptTemplate;\n\n  constructor(fields: HydeRetrieverOptions<V>) {\n    super(fields);\n    this.llm = fields.llm;\n    this.promptTemplate =\n      typeof fields.promptTemplate === \"string\"\n        ? getPromptTemplateFromKey(fields.promptTemplate)\n        : fields.promptTemplate;\n    if (this.promptTemplate) {\n      const { inputVariables } = this.promptTemplate;\n      if (inputVariables.length !== 1 && inputVariables[0] !== \"question\") {\n        throw new Error(\n          `Prompt template must accept a single input variable 'question'. Invalid input variables for prompt template: ${inputVariables}`\n        );\n      }\n    }\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    let value: BasePromptValue = new StringPromptValue(query);\n\n    // Use a custom template if provided\n    if (this.promptTemplate) {\n      value = await this.promptTemplate.formatPromptValue({ question: query });\n    }\n\n    // Get a hypothetical answer from the LLM\n    const res = await this.llm.generatePrompt([value]);\n    const answer = res.generations[0][0].text;\n\n    // Retrieve relevant documents based on the hypothetical answer\n    const results = await this.vectorStore.similaritySearch(\n      answer,\n      this.k,\n      this.filter,\n      runManager?.getChild(\"vectorstore\")\n    );\n\n    return results;\n  }\n}\n\n/**\n * Returns a BasePromptTemplate instance based on a given PromptKey.\n */\nexport function getPromptTemplateFromKey(key: PromptKey): BasePromptTemplate {\n  let template: string;\n\n  switch (key) {\n    case \"websearch\":\n      template = `Please write a passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"scifact\":\n      template = `Please write a scientific paper passage to support/refute the claim\nClaim: {question}\nPassage:`;\n      break;\n    case \"arguana\":\n      template = `Please write a counter argument for the passage\nPassage: {question}\nCounter Argument:`;\n      break;\n    case \"trec-covid\":\n      template = `Please write a scientific paper passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"fiqa\":\n      template = `Please write a financial article passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"dbpedia-entity\":\n      template = `Please write a passage to answer the question.\nQuestion: {question}\nPassage:`;\n      break;\n    case \"trec-news\":\n      template = `Please write a news passage about the topic.\nTopic: {question}\nPassage:`;\n      break;\n    case \"mr-tydi\":\n      template = `Please write a passage in Swahili/Korean/Japanese/Bengali to answer the question in detail.\nQuestion: {question}\nPassage:`;\n      break;\n    default:\n      throw new Error(`Invalid prompt key: ${key}`);\n  }\n\n  return PromptTemplate.fromTemplate(template);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+DA,IAAa,gBAAb,cAEUA,6BAAAA,qBAAwB;CAChC,OAAO,UAAU;AACf,SAAO;;CAGT,IAAI,eAAyB;AAC3B,SAAO;GAAC;GAAa;GAAc;GAAO;;CAG5C;CAEA;CAEA,YAAY,QAAiC;AAC3C,QAAM,OAAO;AACb,OAAK,MAAM,OAAO;AAClB,OAAK,iBACH,OAAO,OAAO,mBAAmB,WAC7B,yBAAyB,OAAO,eAAe,GAC/C,OAAO;AACb,MAAI,KAAK,gBAAgB;GACvB,MAAM,EAAE,mBAAmB,KAAK;AAChC,OAAI,eAAe,WAAW,KAAK,eAAe,OAAO,WACvD,OAAM,IAAI,MACR,gHAAgH,iBACjH;;;CAKP,MAAM,sBACJ,OACA,YACqB;EACrB,IAAI,QAAyB,IAAIC,8BAAAA,kBAAkB,MAAM;AAGzD,MAAI,KAAK,eACP,SAAQ,MAAM,KAAK,eAAe,kBAAkB,EAAE,UAAU,OAAO,CAAC;EAK1E,MAAM,UADM,MAAM,KAAK,IAAI,eAAe,CAAC,MAAM,CAAC,EAC/B,YAAY,GAAG,GAAG;AAUrC,SAPgB,MAAM,KAAK,YAAY,iBACrC,QACA,KAAK,GACL,KAAK,QACL,YAAY,SAAS,cAAc,CACpC;;;;;;AASL,SAAgB,yBAAyB,KAAoC;CAC3E,IAAI;AAEJ,SAAQ,KAAR;EACE,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,QACE,OAAM,IAAI,MAAM,uBAAuB,MAAM;;AAGjD,QAAOC,wBAAAA,eAAe,aAAa,SAAS"}

package/dist/retrievers/hyde.js

@@ -2,7 +2,6 @@ import { __exportAll } from "../_virtual/_rolldown/runtime.js";
 import { PromptTemplate } from "@langchain/core/prompts";
 import { StringPromptValue } from "@langchain/core/prompt_values";
 import { VectorStoreRetriever } from "@langchain/core/vectorstores";
-
 //#region src/retrievers/hyde.ts
 var hyde_exports = /* @__PURE__ */ __exportAll({
 	HydeRetriever: () => HydeRetriever,
@@ -112,7 +111,7 @@ Passage:`;
 	}
 	return PromptTemplate.fromTemplate(template);
 }
-
 //#endregion
 export { HydeRetriever, getPromptTemplateFromKey, hyde_exports };
+
 //# sourceMappingURL=hyde.js.map

package/dist/retrievers/hyde.js.map

@@ -1 +1 @@
-{"version":3,"file":"hyde.js","names":[],"sources":["../../src/retrievers/hyde.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { Document } from \"@langchain/core/documents\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport {\n  StringPromptValue,\n  BasePromptValue,\n} from \"@langchain/core/prompt_values\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\n/**\n * A string that corresponds to a specific prompt template.\n */\nexport type PromptKey =\n  | \"websearch\"\n  | \"scifact\"\n  | \"arguana\"\n  | \"trec-covid\"\n  | \"fiqa\"\n  | \"dbpedia-entity\"\n  | \"trec-news\"\n  | \"mr-tydi\";\n\n/**\n * Options for the HydeRetriever class, which includes a BaseLanguageModel\n * instance, a VectorStore instance, and an optional promptTemplate which\n * can either be a BasePromptTemplate instance or a PromptKey.\n */\nexport type HydeRetrieverOptions<V extends VectorStore> =\n  VectorStoreRetrieverInput<V> & {\n    llm: BaseLanguageModelInterface;\n    promptTemplate?: BasePromptTemplate | PromptKey;\n  };\n\n/**\n * A class for retrieving relevant documents based on a given query. It\n * extends the VectorStoreRetriever class and uses a BaseLanguageModel to\n * generate a hypothetical answer to the query, which is then used to\n * retrieve relevant documents.\n * @example\n * ```typescript\n * const retriever = new HydeRetriever({\n *   vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *   llm: new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *   k: 1,\n * });\n * await vectorStore.addDocuments(\n *   [\n *     \"My name is John.\",\n *     \"My name is Bob.\",\n *     \"My favourite food is pizza.\",\n *     \"My favourite food is pasta.\",\n *   ].map((pageContent) => new Document({ pageContent })),\n * );\n * const results = await retriever.invoke(\n *   \"What is my favourite food?\",\n * );\n * ```\n */\nexport class HydeRetriever<\n  V extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<V> {\n  static lc_name() {\n    return \"HydeRetriever\";\n  }\n\n  get lc_namespace(): string[] {\n    return [\"langchain\", \"retrievers\", \"hyde\"];\n  }\n\n  llm: BaseLanguageModelInterface;\n\n  promptTemplate?: BasePromptTemplate;\n\n  constructor(fields: HydeRetrieverOptions<V>) {\n    super(fields);\n    this.llm = fields.llm;\n    this.promptTemplate =\n      typeof fields.promptTemplate === \"string\"\n        ? getPromptTemplateFromKey(fields.promptTemplate)\n        : fields.promptTemplate;\n    if (this.promptTemplate) {\n      const { inputVariables } = this.promptTemplate;\n      if (inputVariables.length !== 1 && inputVariables[0] !== \"question\") {\n        throw new Error(\n          `Prompt template must accept a single input variable 'question'. Invalid input variables for prompt template: ${inputVariables}`\n        );\n      }\n    }\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    let value: BasePromptValue = new StringPromptValue(query);\n\n    // Use a custom template if provided\n    if (this.promptTemplate) {\n      value = await this.promptTemplate.formatPromptValue({ question: query });\n    }\n\n    // Get a hypothetical answer from the LLM\n    const res = await this.llm.generatePrompt([value]);\n    const answer = res.generations[0][0].text;\n\n    // Retrieve relevant documents based on the hypothetical answer\n    const results = await this.vectorStore.similaritySearch(\n      answer,\n      this.k,\n      this.filter,\n      runManager?.getChild(\"vectorstore\")\n    );\n\n    return results;\n  }\n}\n\n/**\n * Returns a BasePromptTemplate instance based on a given PromptKey.\n */\nexport function getPromptTemplateFromKey(key: PromptKey): BasePromptTemplate {\n  let template: string;\n\n  switch (key) {\n    case \"websearch\":\n      template = `Please write a passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"scifact\":\n      template = `Please write a scientific paper passage to support/refute the claim\nClaim: {question}\nPassage:`;\n      break;\n    case \"arguana\":\n      template = `Please write a counter argument for the passage\nPassage: {question}\nCounter Argument:`;\n      break;\n    case \"trec-covid\":\n      template = `Please write a scientific paper passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"fiqa\":\n      template = `Please write a financial article passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"dbpedia-entity\":\n      template = `Please write a passage to answer the question.\nQuestion: {question}\nPassage:`;\n      break;\n    case \"trec-news\":\n      template = `Please write a news passage about the topic.\nTopic: {question}\nPassage:`;\n      break;\n    case \"mr-tydi\":\n      template = `Please write a passage in Swahili/Korean/Japanese/Bengali to answer the question in detail.\nQuestion: {question}\nPassage:`;\n      break;\n    default:\n      throw new Error(`Invalid prompt key: ${key}`);\n  }\n\n  return PromptTemplate.fromTemplate(template);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+DA,IAAa,gBAAb,cAEU,qBAAwB;CAChC,OAAO,UAAU;AACf,SAAO;;CAGT,IAAI,eAAyB;AAC3B,SAAO;GAAC;GAAa;GAAc;GAAO;;CAG5C;CAEA;CAEA,YAAY,QAAiC;AAC3C,QAAM,OAAO;AACb,OAAK,MAAM,OAAO;AAClB,OAAK,iBACH,OAAO,OAAO,mBAAmB,WAC7B,yBAAyB,OAAO,eAAe,GAC/C,OAAO;AACb,MAAI,KAAK,gBAAgB;GACvB,MAAM,EAAE,mBAAmB,KAAK;AAChC,OAAI,eAAe,WAAW,KAAK,eAAe,OAAO,WACvD,OAAM,IAAI,MACR,gHAAgH,iBACjH;;;CAKP,MAAM,sBACJ,OACA,YACqB;EACrB,IAAI,QAAyB,IAAI,kBAAkB,MAAM;AAGzD,MAAI,KAAK,eACP,SAAQ,MAAM,KAAK,eAAe,kBAAkB,EAAE,UAAU,OAAO,CAAC;EAK1E,MAAM,UADM,MAAM,KAAK,IAAI,eAAe,CAAC,MAAM,CAAC,EAC/B,YAAY,GAAG,GAAG;AAUrC,SAPgB,MAAM,KAAK,YAAY,iBACrC,QACA,KAAK,GACL,KAAK,QACL,YAAY,SAAS,cAAc,CACpC;;;;;;AASL,SAAgB,yBAAyB,KAAoC;CAC3E,IAAI;AAEJ,SAAQ,KAAR;EACE,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,QACE,OAAM,IAAI,MAAM,uBAAuB,MAAM;;AAGjD,QAAO,eAAe,aAAa,SAAS"}
+{"version":3,"file":"hyde.js","names":[],"sources":["../../src/retrievers/hyde.ts"],"sourcesContent":["import type { BaseLanguageModelInterface } from \"@langchain/core/language_models/base\";\nimport { Document } from \"@langchain/core/documents\";\nimport { PromptTemplate, BasePromptTemplate } from \"@langchain/core/prompts\";\nimport {\n  StringPromptValue,\n  BasePromptValue,\n} from \"@langchain/core/prompt_values\";\nimport {\n  VectorStore,\n  VectorStoreRetriever,\n  VectorStoreRetrieverInput,\n} from \"@langchain/core/vectorstores\";\nimport { CallbackManagerForRetrieverRun } from \"@langchain/core/callbacks/manager\";\n\n/**\n * A string that corresponds to a specific prompt template.\n */\nexport type PromptKey =\n  | \"websearch\"\n  | \"scifact\"\n  | \"arguana\"\n  | \"trec-covid\"\n  | \"fiqa\"\n  | \"dbpedia-entity\"\n  | \"trec-news\"\n  | \"mr-tydi\";\n\n/**\n * Options for the HydeRetriever class, which includes a BaseLanguageModel\n * instance, a VectorStore instance, and an optional promptTemplate which\n * can either be a BasePromptTemplate instance or a PromptKey.\n */\nexport type HydeRetrieverOptions<V extends VectorStore> =\n  VectorStoreRetrieverInput<V> & {\n    llm: BaseLanguageModelInterface;\n    promptTemplate?: BasePromptTemplate | PromptKey;\n  };\n\n/**\n * A class for retrieving relevant documents based on a given query. It\n * extends the VectorStoreRetriever class and uses a BaseLanguageModel to\n * generate a hypothetical answer to the query, which is then used to\n * retrieve relevant documents.\n * @example\n * ```typescript\n * const retriever = new HydeRetriever({\n *   vectorStore: new MemoryVectorStore(new OpenAIEmbeddings()),\n *   llm: new ChatOpenAI({ model: \"gpt-4o-mini\" }),\n *   k: 1,\n * });\n * await vectorStore.addDocuments(\n *   [\n *     \"My name is John.\",\n *     \"My name is Bob.\",\n *     \"My favourite food is pizza.\",\n *     \"My favourite food is pasta.\",\n *   ].map((pageContent) => new Document({ pageContent })),\n * );\n * const results = await retriever.invoke(\n *   \"What is my favourite food?\",\n * );\n * ```\n */\nexport class HydeRetriever<\n  V extends VectorStore = VectorStore,\n> extends VectorStoreRetriever<V> {\n  static lc_name() {\n    return \"HydeRetriever\";\n  }\n\n  get lc_namespace(): string[] {\n    return [\"langchain\", \"retrievers\", \"hyde\"];\n  }\n\n  llm: BaseLanguageModelInterface;\n\n  promptTemplate?: BasePromptTemplate;\n\n  constructor(fields: HydeRetrieverOptions<V>) {\n    super(fields);\n    this.llm = fields.llm;\n    this.promptTemplate =\n      typeof fields.promptTemplate === \"string\"\n        ? getPromptTemplateFromKey(fields.promptTemplate)\n        : fields.promptTemplate;\n    if (this.promptTemplate) {\n      const { inputVariables } = this.promptTemplate;\n      if (inputVariables.length !== 1 && inputVariables[0] !== \"question\") {\n        throw new Error(\n          `Prompt template must accept a single input variable 'question'. Invalid input variables for prompt template: ${inputVariables}`\n        );\n      }\n    }\n  }\n\n  async _getRelevantDocuments(\n    query: string,\n    runManager?: CallbackManagerForRetrieverRun\n  ): Promise<Document[]> {\n    let value: BasePromptValue = new StringPromptValue(query);\n\n    // Use a custom template if provided\n    if (this.promptTemplate) {\n      value = await this.promptTemplate.formatPromptValue({ question: query });\n    }\n\n    // Get a hypothetical answer from the LLM\n    const res = await this.llm.generatePrompt([value]);\n    const answer = res.generations[0][0].text;\n\n    // Retrieve relevant documents based on the hypothetical answer\n    const results = await this.vectorStore.similaritySearch(\n      answer,\n      this.k,\n      this.filter,\n      runManager?.getChild(\"vectorstore\")\n    );\n\n    return results;\n  }\n}\n\n/**\n * Returns a BasePromptTemplate instance based on a given PromptKey.\n */\nexport function getPromptTemplateFromKey(key: PromptKey): BasePromptTemplate {\n  let template: string;\n\n  switch (key) {\n    case \"websearch\":\n      template = `Please write a passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"scifact\":\n      template = `Please write a scientific paper passage to support/refute the claim\nClaim: {question}\nPassage:`;\n      break;\n    case \"arguana\":\n      template = `Please write a counter argument for the passage\nPassage: {question}\nCounter Argument:`;\n      break;\n    case \"trec-covid\":\n      template = `Please write a scientific paper passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"fiqa\":\n      template = `Please write a financial article passage to answer the question\nQuestion: {question}\nPassage:`;\n      break;\n    case \"dbpedia-entity\":\n      template = `Please write a passage to answer the question.\nQuestion: {question}\nPassage:`;\n      break;\n    case \"trec-news\":\n      template = `Please write a news passage about the topic.\nTopic: {question}\nPassage:`;\n      break;\n    case \"mr-tydi\":\n      template = `Please write a passage in Swahili/Korean/Japanese/Bengali to answer the question in detail.\nQuestion: {question}\nPassage:`;\n      break;\n    default:\n      throw new Error(`Invalid prompt key: ${key}`);\n  }\n\n  return PromptTemplate.fromTemplate(template);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+DA,IAAa,gBAAb,cAEU,qBAAwB;CAChC,OAAO,UAAU;AACf,SAAO;;CAGT,IAAI,eAAyB;AAC3B,SAAO;GAAC;GAAa;GAAc;GAAO;;CAG5C;CAEA;CAEA,YAAY,QAAiC;AAC3C,QAAM,OAAO;AACb,OAAK,MAAM,OAAO;AAClB,OAAK,iBACH,OAAO,OAAO,mBAAmB,WAC7B,yBAAyB,OAAO,eAAe,GAC/C,OAAO;AACb,MAAI,KAAK,gBAAgB;GACvB,MAAM,EAAE,mBAAmB,KAAK;AAChC,OAAI,eAAe,WAAW,KAAK,eAAe,OAAO,WACvD,OAAM,IAAI,MACR,gHAAgH,iBACjH;;;CAKP,MAAM,sBACJ,OACA,YACqB;EACrB,IAAI,QAAyB,IAAI,kBAAkB,MAAM;AAGzD,MAAI,KAAK,eACP,SAAQ,MAAM,KAAK,eAAe,kBAAkB,EAAE,UAAU,OAAO,CAAC;EAK1E,MAAM,UADM,MAAM,KAAK,IAAI,eAAe,CAAC,MAAM,CAAC,EAC/B,YAAY,GAAG,GAAG;AAUrC,SAPgB,MAAM,KAAK,YAAY,iBACrC,QACA,KAAK,GACL,KAAK,QACL,YAAY,SAAS,cAAc,CACpC;;;;;;AASL,SAAgB,yBAAyB,KAAoC;CAC3E,IAAI;AAEJ,SAAQ,KAAR;EACE,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,KAAK;AACH,cAAW;;;AAGX;EACF,QACE,OAAM,IAAI,MAAM,uBAAuB,MAAM;;AAGjD,QAAO,eAAe,aAAa,SAAS"}

package/dist/retrievers/matryoshka_retriever.cjs

@@ -1,8 +1,7 @@
-Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });
-const require_runtime = require('../_virtual/_rolldown/runtime.cjs');
+Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
+const require_runtime = require("../_virtual/_rolldown/runtime.cjs");
 let _langchain_core_utils_math = require("@langchain/core/utils/math");
 let _langchain_core_vectorstores = require("@langchain/core/vectorstores");
-
 //#region src/retrievers/matryoshka_retriever.ts
 var matryoshka_retriever_exports = /* @__PURE__ */ require_runtime.__exportAll({ MatryoshkaRetriever: () => MatryoshkaRetriever });
 /**
@@ -104,13 +103,13 @@ var MatryoshkaRetriever = class extends _langchain_core_vectorstores.VectorStore
 		return this.vectorStore.addDocuments(newDocuments, options);
 	};
 };
-
 //#endregion
 exports.MatryoshkaRetriever = MatryoshkaRetriever;
-Object.defineProperty(exports, 'matryoshka_retriever_exports', {
-  enumerable: true,
-  get: function () {
-    return matryoshka_retriever_exports;
-  }
+Object.defineProperty(exports, "matryoshka_retriever_exports", {
+	enumerable: true,
+	get: function() {
+		return matryoshka_retriever_exports;
+	}
 });
+
 //# sourceMappingURL=matryoshka_retriever.cjs.map

package/dist/retrievers/matryoshka_retriever.cjs.map

@@ -1 +1 @@
-{"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"}
+{"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,6BAAAA,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,uBACE,eACA,cACqB;EACrB,MAAM,kBAAmC,aAAa,KAAK,QACzD,KAAK,MAAM,IAAI,SAAS,KAAK,mBAAmB,CACjD;EACD,IAAI;AAEJ,UAAQ,KAAK,YAAb;GACE,KAAK;AACH,kBAAA,GAAA,2BAAA,kBAA8B,CAAC,cAAc,EAAE,gBAAgB;AAC/D;GACF,KAAK;AACH,kBAAA,GAAA,2BAAA,cAA0B,CAAC,cAAc,EAAE,gBAAgB;AAC3D;GACF,KAAK;AACH,kBAAA,GAAA,2BAAA,mBAA+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,eAAwB,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.js

@@ -1,7 +1,6 @@
 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 = /* @__PURE__ */ __exportAll({ MatryoshkaRetriever: () => MatryoshkaRetriever });
 /**
@@ -103,7 +102,7 @@ var MatryoshkaRetriever = class extends VectorStoreRetriever {
 		return this.vectorStore.addDocuments(newDocuments, options);
 	};
 };
-
 //#endregion
 export { MatryoshkaRetriever, matryoshka_retriever_exports };
+
 //# sourceMappingURL=matryoshka_retriever.js.map

package/dist/retrievers/matryoshka_retriever.js.map

@@ -1 +1 @@
-{"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"}
+{"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,uBACE,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,eAAwB,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

@@ -1,10 +1,9 @@
-Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' });
-const require_runtime = require('../_virtual/_rolldown/runtime.cjs');
-const require_llm_chain = require('../chains/llm_chain.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");
 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 = /* @__PURE__ */ require_runtime.__exportAll({ MultiQueryRetriever: () => MultiQueryRetriever });
 var LineListOutputParser = class extends _langchain_core_output_parsers.BaseOutputParser {
@@ -134,13 +133,13 @@ var MultiQueryRetriever = class extends _langchain_core_retrievers.BaseRetriever
 		return outputDocs;
 	}
 };
-
 //#endregion
 exports.MultiQueryRetriever = MultiQueryRetriever;
-Object.defineProperty(exports, 'multi_query_exports', {
-  enumerable: true,
-  get: function () {
-    return multi_query_exports;
-  }
+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

@@ -1 +1 @@
-{"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"}
+{"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,+BAAAA,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,IAAIC,wBAAAA,eAAe;CAC9D,gBAAgB,CAAC,YAAY,aAAa;CAC1C,UAAU;;;;;;;;;;;;;;;;CAgBX,CAAC;;;;;;;;;;;;;;AAyBF,IAAa,sBAAb,cAAyCC,2BAAAA,cAAc;CACrD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAa;CAExD;CAEA;CAEA,aAAqB;CAErB,YAAoB;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,kBAAAA,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,aAAqB,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.js

@@ -3,7 +3,6 @@ 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 = /* @__PURE__ */ __exportAll({ MultiQueryRetriever: () => MultiQueryRetriever });
 var LineListOutputParser = class extends BaseOutputParser {
@@ -133,7 +132,7 @@ var MultiQueryRetriever = class extends BaseRetriever {
 		return outputDocs;
 	}
 };
-
 //#endregion
 export { MultiQueryRetriever, multi_query_exports };
+
 //# sourceMappingURL=multi_query.js.map

package/dist/retrievers/multi_query.js.map

@@ -1 +1 @@
-{"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"}
+{"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;CAEA;CAEA,aAAqB;CAErB,YAAoB;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,aAAqB,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

@@ -1,8 +1,7 @@
-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');
+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");
 let _langchain_core_retrievers = require("@langchain/core/retrievers");
-
 //#region src/retrievers/multi_vector.ts
 var multi_vector_exports = /* @__PURE__ */ require_runtime.__exportAll({ MultiVectorRetriever: () => MultiVectorRetriever });
 /**
@@ -54,13 +53,13 @@ var MultiVectorRetriever = class extends _langchain_core_retrievers.BaseRetrieve
 		return (await this.docstore.mget(ids)).filter((doc) => doc !== void 0).slice(0, this.parentK);
 	}
 };
-
 //#endregion
 exports.MultiVectorRetriever = MultiVectorRetriever;
-Object.defineProperty(exports, 'multi_vector_exports', {
-  enumerable: true,
-  get: function () {
-    return multi_vector_exports;
-  }
+Object.defineProperty(exports, "multi_vector_exports", {
+	enumerable: true,
+	get: function() {
+		return multi_vector_exports;
+	}
 });
+
 //# sourceMappingURL=multi_vector.cjs.map

package/dist/retrievers/multi_vector.cjs.map

@@ -1 +1 @@
-{"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"}
+{"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,2BAAAA,cAAc;CACtD,OAAO,UAAU;AACf,SAAO;;CAGT,eAAe;EAAC;EAAa;EAAc;EAAe;CAE1D;CAEA;CAEA;CAEA;CAEA;CAEA,YAAY,MAAiC;AAC3C,QAAM,KAAK;AACX,OAAK,cAAc,KAAK;AACxB,MAAI,KAAK,UACP,MAAK,WAAWC,+BAAAA,iCAAiC,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,KAAA,EAAU,CAClC,MAAM,GAAG,KAAK,QAAQ"}

package/dist/retrievers/multi_vector.js

@@ -1,7 +1,6 @@
 import { __exportAll } from "../_virtual/_rolldown/runtime.js";
 import { createDocumentStoreFromByteStore } from "../storage/encoder_backed.js";
 import { BaseRetriever } from "@langchain/core/retrievers";
-
 //#region src/retrievers/multi_vector.ts
 var multi_vector_exports = /* @__PURE__ */ __exportAll({ MultiVectorRetriever: () => MultiVectorRetriever });
 /**
@@ -53,7 +52,7 @@ var MultiVectorRetriever = class extends BaseRetriever {
 		return (await this.docstore.mget(ids)).filter((doc) => doc !== void 0).slice(0, this.parentK);
 	}
 };
-
 //#endregion
 export { MultiVectorRetriever, multi_vector_exports };
+
 //# sourceMappingURL=multi_vector.js.map