langchain 0.0.173 → 0.0.175

package/dist/vectorstores/convex.d.ts

@@ -0,0 +1,113 @@
+import { DocumentByInfo, FieldPaths, FilterExpression, FunctionReference, GenericActionCtx, GenericDataModel, GenericTableInfo, NamedTableInfo, NamedVectorIndex, TableNamesInDataModel, VectorFilterBuilder, VectorIndexNames } from "convex/server";
+import { Document } from "../document.js";
+import { Embeddings } from "../embeddings/base.js";
+import { VectorStore } from "./base.js";
+/**
+ * Type that defines the config required to initialize the
+ * ConvexVectorStore class. It includes the table name,
+ * index name, text field name, and embedding field name.
+ */
+export type ConvexVectorStoreConfig<DataModel extends GenericDataModel, TableName extends TableNamesInDataModel<DataModel>, IndexName extends VectorIndexNames<NamedTableInfo<DataModel, TableName>>, TextFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, EmbeddingFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, MetadataFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, InsertMutation extends FunctionReference<"mutation", "internal", {
+    table: string;
+    document: object;
+}>, GetQuery extends FunctionReference<"query", "internal", {
+    id: string;
+}, object | null>> = {
+    readonly ctx: GenericActionCtx<DataModel>;
+    readonly table?: TableName;
+    readonly index?: IndexName;
+    readonly textField?: TextFieldName;
+    readonly embeddingField?: EmbeddingFieldName;
+    readonly metadataField?: MetadataFieldName;
+    readonly insert?: InsertMutation;
+    readonly get?: GetQuery;
+};
+/**
+ * Class that is a wrapper around Convex storage and vector search. It is used
+ * to insert embeddings in Convex documents with a vector search index,
+ * and perform a vector search on them.
+ *
+ * ConvexVectorStore does NOT implement maxMarginalRelevanceSearch.
+ */
+export declare class ConvexVectorStore<DataModel extends GenericDataModel, TableName extends TableNamesInDataModel<DataModel>, IndexName extends VectorIndexNames<NamedTableInfo<DataModel, TableName>>, TextFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, EmbeddingFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, MetadataFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, InsertMutation extends FunctionReference<"mutation", "internal", {
+    table: string;
+    document: object;
+}>, GetQuery extends FunctionReference<"query", "internal", {
+    id: string;
+}, object | null>> extends VectorStore {
+    /**
+     * Type that defines the filter used in the
+     * similaritySearchVectorWithScore and maxMarginalRelevanceSearch methods.
+     * It includes limit, filter and a flag to include embeddings.
+     */
+    FilterType: {
+        filter?: (q: VectorFilterBuilder<DocumentByInfo<GenericTableInfo>, NamedVectorIndex<NamedTableInfo<DataModel, TableName>, IndexName>>) => FilterExpression<boolean>;
+        includeEmbeddings?: boolean;
+    };
+    private readonly ctx;
+    private readonly table;
+    private readonly index;
+    private readonly textField;
+    private readonly embeddingField;
+    private readonly metadataField;
+    private readonly insert;
+    private readonly get;
+    _vectorstoreType(): string;
+    constructor(embeddings: Embeddings, config: ConvexVectorStoreConfig<DataModel, TableName, IndexName, TextFieldName, EmbeddingFieldName, MetadataFieldName, InsertMutation, GetQuery>);
+    /**
+     * Add vectors and their corresponding documents to the Convex table.
+     * @param vectors Vectors to be added.
+     * @param documents Corresponding documents to be added.
+     * @returns Promise that resolves when the vectors and documents have been added.
+     */
+    addVectors(vectors: number[][], documents: Document[]): Promise<void>;
+    /**
+     * Add documents to the Convex table. It first converts
+     * the documents to vectors using the embeddings and then calls the
+     * addVectors method.
+     * @param documents Documents to be added.
+     * @returns Promise that resolves when the documents have been added.
+     */
+    addDocuments(documents: Document[]): Promise<void>;
+    /**
+     * Similarity search on the vectors stored in the
+     * Convex table. It returns a list of documents and their
+     * corresponding similarity scores.
+     * @param query Query vector for the similarity search.
+     * @param k Number of nearest neighbors to return.
+     * @param filter Optional filter to be applied.
+     * @returns Promise that resolves to a list of documents and their corresponding similarity scores.
+     */
+    similaritySearchVectorWithScore(query: number[], k: number, filter?: this["FilterType"]): Promise<[Document, number][]>;
+    /**
+     * Static method to create an instance of ConvexVectorStore from a
+     * list of texts. It first converts the texts to vectors and then adds
+     * them to the Convex table.
+     * @param texts List of texts to be converted to vectors.
+     * @param metadatas Metadata for the texts.
+     * @param embeddings Embeddings to be used for conversion.
+     * @param dbConfig Database configuration for Convex.
+     * @returns Promise that resolves to a new instance of ConvexVectorStore.
+     */
+    static fromTexts<DataModel extends GenericDataModel, TableName extends TableNamesInDataModel<DataModel>, IndexName extends VectorIndexNames<NamedTableInfo<DataModel, TableName>>, TextFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, EmbeddingFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, MetadataFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, InsertMutation extends FunctionReference<"mutation", "internal", {
+        table: string;
+        document: object;
+    }>, GetQuery extends FunctionReference<"query", "internal", {
+        id: string;
+    }, object | null>>(texts: string[], metadatas: object[] | object, embeddings: Embeddings, dbConfig: ConvexVectorStoreConfig<DataModel, TableName, IndexName, TextFieldName, EmbeddingFieldName, MetadataFieldName, InsertMutation, GetQuery>): Promise<ConvexVectorStore<DataModel, TableName, IndexName, TextFieldName, EmbeddingFieldName, MetadataFieldName, InsertMutation, GetQuery>>;
+    /**
+     * Static method to create an instance of ConvexVectorStore from a
+     * list of documents. It first converts the documents to vectors and then
+     * adds them to the Convex table.
+     * @param docs List of documents to be converted to vectors.
+     * @param embeddings Embeddings to be used for conversion.
+     * @param dbConfig Database configuration for Convex.
+     * @returns Promise that resolves to a new instance of ConvexVectorStore.
+     */
+    static fromDocuments<DataModel extends GenericDataModel, TableName extends TableNamesInDataModel<DataModel>, IndexName extends VectorIndexNames<NamedTableInfo<DataModel, TableName>>, TextFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, EmbeddingFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, MetadataFieldName extends FieldPaths<NamedTableInfo<DataModel, TableName>>, InsertMutation extends FunctionReference<"mutation", "internal", {
+        table: string;
+        document: object;
+    }>, GetQuery extends FunctionReference<"query", "internal", {
+        id: string;
+    }, object | null>>(docs: Document[], embeddings: Embeddings, dbConfig: ConvexVectorStoreConfig<DataModel, TableName, IndexName, TextFieldName, EmbeddingFieldName, MetadataFieldName, InsertMutation, GetQuery>): Promise<ConvexVectorStore<DataModel, TableName, IndexName, TextFieldName, EmbeddingFieldName, MetadataFieldName, InsertMutation, GetQuery>>;
+}

package/dist/vectorstores/convex.js

@@ -0,0 +1,173 @@
+// eslint-disable-next-line import/no-extraneous-dependencies
+import { makeFunctionReference, } from "convex/server";
+import { Document } from "../document.js";
+import { VectorStore } from "./base.js";
+/**
+ * Class that is a wrapper around Convex storage and vector search. It is used
+ * to insert embeddings in Convex documents with a vector search index,
+ * and perform a vector search on them.
+ *
+ * ConvexVectorStore does NOT implement maxMarginalRelevanceSearch.
+ */
+export class ConvexVectorStore extends VectorStore {
+    _vectorstoreType() {
+        return "convex";
+    }
+    constructor(embeddings, config) {
+        super(embeddings, config);
+        Object.defineProperty(this, "ctx", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "table", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "index", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "textField", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "embeddingField", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "metadataField", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "insert", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        Object.defineProperty(this, "get", {
+            enumerable: true,
+            configurable: true,
+            writable: true,
+            value: void 0
+        });
+        this.ctx = config.ctx;
+        this.table = config.table ?? "documents";
+        this.index = config.index ?? "byEmbedding";
+        this.textField = config.textField ?? "text";
+        this.embeddingField =
+            config.embeddingField ?? "embedding";
+        this.metadataField =
+            config.metadataField ?? "metadata";
+        this.insert =
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            config.insert ?? makeFunctionReference("langchain/db:insert");
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        this.get = config.get ?? makeFunctionReference("langchain/db:get");
+    }
+    /**
+     * Add vectors and their corresponding documents to the Convex table.
+     * @param vectors Vectors to be added.
+     * @param documents Corresponding documents to be added.
+     * @returns Promise that resolves when the vectors and documents have been added.
+     */
+    async addVectors(vectors, documents) {
+        const convexDocuments = vectors.map((embedding, idx) => ({
+            [this.textField]: documents[idx].pageContent,
+            [this.embeddingField]: embedding,
+            [this.metadataField]: documents[idx].metadata,
+        }));
+        // TODO: Remove chunking when Convex handles the concurrent requests correctly
+        const PAGE_SIZE = 16;
+        for (let i = 0; i < convexDocuments.length; i += PAGE_SIZE) {
+            await Promise.all(convexDocuments.slice(i, i + PAGE_SIZE).map((document) => this.ctx.runMutation(this.insert, {
+                table: this.table,
+                document,
+                // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            })));
+        }
+    }
+    /**
+     * Add documents to the Convex table. It first converts
+     * the documents to vectors using the embeddings and then calls the
+     * addVectors method.
+     * @param documents Documents to be added.
+     * @returns Promise that resolves when the documents have been added.
+     */
+    async addDocuments(documents) {
+        const texts = documents.map(({ pageContent }) => pageContent);
+        return this.addVectors(await this.embeddings.embedDocuments(texts), documents);
+    }
+    /**
+     * Similarity search on the vectors stored in the
+     * Convex table. It returns a list of documents and their
+     * corresponding similarity scores.
+     * @param query Query vector for the similarity search.
+     * @param k Number of nearest neighbors to return.
+     * @param filter Optional filter to be applied.
+     * @returns Promise that resolves to a list of documents and their corresponding similarity scores.
+     */
+    async similaritySearchVectorWithScore(query, k, filter) {
+        const idsAndScores = await this.ctx.vectorSearch(this.table, this.index, {
+            vector: query,
+            limit: k,
+            filter: filter?.filter,
+        });
+        const documents = await Promise.all(idsAndScores.map(({ _id }) => 
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        this.ctx.runQuery(this.get, { id: _id })));
+        return documents.map(({ [this.textField]: text, [this.embeddingField]: embedding, [this.metadataField]: metadata, }, idx) => [
+            new Document({
+                pageContent: text,
+                metadata: {
+                    ...metadata,
+                    ...(filter?.includeEmbeddings ? { embedding } : null),
+                },
+            }),
+            idsAndScores[idx]._score,
+        ]);
+    }
+    /**
+     * Static method to create an instance of ConvexVectorStore from a
+     * list of texts. It first converts the texts to vectors and then adds
+     * them to the Convex table.
+     * @param texts List of texts to be converted to vectors.
+     * @param metadatas Metadata for the texts.
+     * @param embeddings Embeddings to be used for conversion.
+     * @param dbConfig Database configuration for Convex.
+     * @returns Promise that resolves to a new instance of ConvexVectorStore.
+     */
+    static async fromTexts(texts, metadatas, embeddings, dbConfig) {
+        const docs = texts.map((text, i) => new Document({
+            pageContent: text,
+            metadata: Array.isArray(metadatas) ? metadatas[i] : metadatas,
+        }));
+        return ConvexVectorStore.fromDocuments(docs, embeddings, dbConfig);
+    }
+    /**
+     * Static method to create an instance of ConvexVectorStore from a
+     * list of documents. It first converts the documents to vectors and then
+     * adds them to the Convex table.
+     * @param docs List of documents to be converted to vectors.
+     * @param embeddings Embeddings to be used for conversion.
+     * @param dbConfig Database configuration for Convex.
+     * @returns Promise that resolves to a new instance of ConvexVectorStore.
+     */
+    static async fromDocuments(docs, embeddings, dbConfig) {
+        const instance = new this(embeddings, dbConfig);
+        await instance.addDocuments(docs);
+        return instance;
+    }
+}

package/dist/vectorstores/milvus.cjs

@@ -577,11 +577,13 @@ function genCollectionName() {
 }
 function getTextFieldMaxLength(documents) {
     let textMaxLength = 0;
+    const textEncoder = new TextEncoder();
     // eslint-disable-next-line no-plusplus
     for (let i = 0; i < documents.length; i++) {
         const text = documents[i].pageContent;
-        if (text.length > textMaxLength) {
-            textMaxLength = text.length;
+        const textLengthInBytes = textEncoder.encode(text).length;
+        if (textLengthInBytes > textMaxLength) {
+            textMaxLength = textLengthInBytes;
         }
     }
     return textMaxLength;

package/dist/vectorstores/milvus.js

@@ -550,11 +550,13 @@ function genCollectionName() {
 }
 function getTextFieldMaxLength(documents) {
     let textMaxLength = 0;
+    const textEncoder = new TextEncoder();
     // eslint-disable-next-line no-plusplus
     for (let i = 0; i < documents.length; i++) {
         const text = documents[i].pageContent;
-        if (text.length > textMaxLength) {
-            textMaxLength = text.length;
+        const textLengthInBytes = textEncoder.encode(text).length;
+        if (textLengthInBytes > textMaxLength) {
+            textMaxLength = textLengthInBytes;
         }
     }
     return textMaxLength;

package/dist/vectorstores/pinecone.cjs

@@ -33,6 +33,7 @@ const flat_1 = __importDefault(require("flat"));
 const base_js_1 = require("./base.cjs");
 const document_js_1 = require("../document.cjs");
 const async_caller_js_1 = require("../util/async_caller.cjs");
+const math_js_1 = require("../util/math.cjs");
 /**
  * Class that extends the VectorStore class and provides methods to
  * interact with the Pinecone vector database.
@@ -166,15 +167,7 @@ class PineconeStore extends base_js_1.VectorStore {
             throw new Error("Either ids or delete_all must be provided.");
         }
     }
-    /**
-     * Method that performs a similarity search in the Pinecone database and
-     * returns the results along with their scores.
-     * @param query Query vector for the similarity search.
-     * @param k Number of top results to return.
-     * @param filter Optional filter to apply to the search.
-     * @returns Promise that resolves with an array of documents and their scores.
-     */
-    async similaritySearchVectorWithScore(query, k, filter) {
+    async _runPineconeQuery(query, k, filter, options) {
         if (filter && this.filter) {
             throw new Error("cannot provide both `filter` and `this.filter`");
         }
@@ -185,7 +178,20 @@ class PineconeStore extends base_js_1.VectorStore {
             topK: k,
             vector: query,
             filter: _filter,
+            ...options,
         });
+        return results;
+    }
+    /**
+     * Method that performs a similarity search in the Pinecone database and
+     * returns the results along with their scores.
+     * @param query Query vector for the similarity search.
+     * @param k Number of top results to return.
+     * @param filter Optional filter to apply to the search.
+     * @returns Promise that resolves with an array of documents and their scores.
+     */
+    async similaritySearchVectorWithScore(query, k, filter) {
+        const results = await this._runPineconeQuery(query, k, filter);
         const result = [];
         if (results.matches) {
             for (const res of results.matches) {
@@ -198,6 +204,37 @@ class PineconeStore extends base_js_1.VectorStore {
         }
         return result;
     }
+    /**
+     * Return documents selected using the maximal marginal relevance.
+     * Maximal marginal relevance optimizes for similarity to the query AND diversity
+     * among selected documents.
+     *
+     * @param {string} query - Text to look up documents similar to.
+     * @param {number} options.k - Number of documents to return.
+     * @param {number} options.fetchK=20 - Number of documents to fetch before passing to the MMR algorithm.
+     * @param {number} options.lambda=0.5 - Number between 0 and 1 that determines the degree of diversity among the results,
+     *                 where 0 corresponds to maximum diversity and 1 to minimum diversity.
+     * @param {PineconeMetadata} options.filter - Optional filter to apply to the search.
+     *
+     * @returns {Promise<Document[]>} - List of documents selected by maximal marginal relevance.
+     */
+    async maxMarginalRelevanceSearch(query, options) {
+        const queryEmbedding = await this.embeddings.embedQuery(query);
+        const results = await this._runPineconeQuery(queryEmbedding, options.fetchK ?? 20, options.filter, { includeValues: true });
+        const matches = results?.matches ?? [];
+        const embeddingList = matches.map((match) => match.values);
+        const mmrIndexes = (0, math_js_1.maximalMarginalRelevance)(queryEmbedding, embeddingList, options.lambda, options.k);
+        const topMmrMatches = mmrIndexes.map((idx) => matches[idx]);
+        const finalResult = [];
+        for (const res of topMmrMatches) {
+            const { [this.textKey]: pageContent, ...metadata } = (res.metadata ??
+                {});
+            if (res.score) {
+                finalResult.push(new document_js_1.Document({ metadata, pageContent }));
+            }
+        }
+        return finalResult;
+    }
     /**
      * Static method that creates a new instance of the PineconeStore class
      * from texts.

package/dist/vectorstores/pinecone.d.ts

@@ -1,5 +1,5 @@
-import { Index as PineconeIndex } from "@pinecone-database/pinecone";
-import { VectorStore } from "./base.js";
+import { RecordMetadata, Index as PineconeIndex } from "@pinecone-database/pinecone";
+import { MaxMarginalRelevanceSearchOptions, VectorStore } from "./base.js";
 import { Embeddings } from "../embeddings/base.js";
 import { Document } from "../document.js";
 import { AsyncCaller } from "../util/async_caller.js";
@@ -57,6 +57,9 @@ export declare class PineconeStore extends VectorStore {
      * @returns Promise that resolves when the delete operation is complete.
      */
     delete(params: PineconeDeleteParams): Promise<void>;
+    protected _runPineconeQuery(query: number[], k: number, filter?: PineconeMetadata, options?: {
+        includeValues: boolean;
+    }): Promise<import("@pinecone-database/pinecone").QueryResponse<RecordMetadata>>;
     /**
      * Method that performs a similarity search in the Pinecone database and
      * returns the results along with their scores.
@@ -66,6 +69,21 @@ export declare class PineconeStore extends VectorStore {
      * @returns Promise that resolves with an array of documents and their scores.
      */
     similaritySearchVectorWithScore(query: number[], k: number, filter?: PineconeMetadata): Promise<[Document, number][]>;
+    /**
+     * Return documents selected using the maximal marginal relevance.
+     * Maximal marginal relevance optimizes for similarity to the query AND diversity
+     * among selected documents.
+     *
+     * @param {string} query - Text to look up documents similar to.
+     * @param {number} options.k - Number of documents to return.
+     * @param {number} options.fetchK=20 - Number of documents to fetch before passing to the MMR algorithm.
+     * @param {number} options.lambda=0.5 - Number between 0 and 1 that determines the degree of diversity among the results,
+     *                 where 0 corresponds to maximum diversity and 1 to minimum diversity.
+     * @param {PineconeMetadata} options.filter - Optional filter to apply to the search.
+     *
+     * @returns {Promise<Document[]>} - List of documents selected by maximal marginal relevance.
+     */
+    maxMarginalRelevanceSearch(query: string, options: MaxMarginalRelevanceSearchOptions<this["FilterType"]>): Promise<Document[]>;
     /**
      * Static method that creates a new instance of the PineconeStore class
      * from texts.

package/dist/vectorstores/pinecone.js

@@ -4,6 +4,7 @@ import flatten from "flat";
 import { VectorStore } from "./base.js";
 import { Document } from "../document.js";
 import { AsyncCaller } from "../util/async_caller.js";
+import { maximalMarginalRelevance } from "../util/math.js";
 /**
  * Class that extends the VectorStore class and provides methods to
  * interact with the Pinecone vector database.
@@ -137,15 +138,7 @@ export class PineconeStore extends VectorStore {
             throw new Error("Either ids or delete_all must be provided.");
         }
     }
-    /**
-     * Method that performs a similarity search in the Pinecone database and
-     * returns the results along with their scores.
-     * @param query Query vector for the similarity search.
-     * @param k Number of top results to return.
-     * @param filter Optional filter to apply to the search.
-     * @returns Promise that resolves with an array of documents and their scores.
-     */
-    async similaritySearchVectorWithScore(query, k, filter) {
+    async _runPineconeQuery(query, k, filter, options) {
         if (filter && this.filter) {
             throw new Error("cannot provide both `filter` and `this.filter`");
         }
@@ -156,7 +149,20 @@ export class PineconeStore extends VectorStore {
             topK: k,
             vector: query,
             filter: _filter,
+            ...options,
         });
+        return results;
+    }
+    /**
+     * Method that performs a similarity search in the Pinecone database and
+     * returns the results along with their scores.
+     * @param query Query vector for the similarity search.
+     * @param k Number of top results to return.
+     * @param filter Optional filter to apply to the search.
+     * @returns Promise that resolves with an array of documents and their scores.
+     */
+    async similaritySearchVectorWithScore(query, k, filter) {
+        const results = await this._runPineconeQuery(query, k, filter);
         const result = [];
         if (results.matches) {
             for (const res of results.matches) {
@@ -169,6 +175,37 @@ export class PineconeStore extends VectorStore {
         }
         return result;
     }
+    /**
+     * Return documents selected using the maximal marginal relevance.
+     * Maximal marginal relevance optimizes for similarity to the query AND diversity
+     * among selected documents.
+     *
+     * @param {string} query - Text to look up documents similar to.
+     * @param {number} options.k - Number of documents to return.
+     * @param {number} options.fetchK=20 - Number of documents to fetch before passing to the MMR algorithm.
+     * @param {number} options.lambda=0.5 - Number between 0 and 1 that determines the degree of diversity among the results,
+     *                 where 0 corresponds to maximum diversity and 1 to minimum diversity.
+     * @param {PineconeMetadata} options.filter - Optional filter to apply to the search.
+     *
+     * @returns {Promise<Document[]>} - List of documents selected by maximal marginal relevance.
+     */
+    async maxMarginalRelevanceSearch(query, options) {
+        const queryEmbedding = await this.embeddings.embedQuery(query);
+        const results = await this._runPineconeQuery(queryEmbedding, options.fetchK ?? 20, options.filter, { includeValues: true });
+        const matches = results?.matches ?? [];
+        const embeddingList = matches.map((match) => match.values);
+        const mmrIndexes = maximalMarginalRelevance(queryEmbedding, embeddingList, options.lambda, options.k);
+        const topMmrMatches = mmrIndexes.map((idx) => matches[idx]);
+        const finalResult = [];
+        for (const res of topMmrMatches) {
+            const { [this.textKey]: pageContent, ...metadata } = (res.metadata ??
+                {});
+            if (res.score) {
+                finalResult.push(new Document({ metadata, pageContent }));
+            }
+        }
+        return finalResult;
+    }
     /**
      * Static method that creates a new instance of the PineconeStore class
      * from texts.

package/dist/vectorstores/vercel_postgres.cjs

@@ -206,14 +206,36 @@ class VercelPostgres extends base_js_1.VectorStore {
      */
     async similaritySearchVectorWithScore(query, k, filter) {
         const embeddingString = `[${query.join(",")}]`;
-        const _filter = filter ?? "{}";
+        const _filter = filter ?? {};
+        const whereClauses = [];
+        const values = [embeddingString, k];
+        let paramCount = values.length;
+        for (const [key, value] of Object.entries(_filter)) {
+            if (typeof value === "object" && value !== null) {
+                const currentParamCount = paramCount;
+                const placeholders = value.in
+                    .map((_, index) => `$${currentParamCount + index + 1}`)
+                    .join(",");
+                whereClauses.push(`${this.metadataColumnName}->>'${key}' IN (${placeholders})`);
+                values.push(...value.in);
+                paramCount += value.in.length;
+            }
+            else {
+                paramCount += 1;
+                whereClauses.push(`${this.metadataColumnName}->>'${key}' = $${paramCount}`);
+                values.push(value);
+            }
+        }
+        const whereClause = whereClauses.length
+            ? `WHERE ${whereClauses.join(" AND ")}`
+            : "";
         const queryString = `
-      SELECT *, ${this.vectorColumnName} <=> $1 as "_distance"
-      FROM ${this.tableName}
-      WHERE ${this.metadataColumnName} @> $2
-      ORDER BY "_distance" ASC
-      LIMIT $3;`;
-        const documents = (await this.client.query(queryString, [embeddingString, _filter, k])).rows;
+            SELECT *, ${this.vectorColumnName} <=> $1 as "_distance"
+            FROM ${this.tableName}
+            ${whereClause}
+            ORDER BY "_distance" ASC
+            LIMIT $2;`;
+        const documents = (await this.client.query(queryString, values)).rows;
         const results = [];
         for (const doc of documents) {
             if (doc._distance != null && doc[this.contentColumnName] != null) {

package/dist/vectorstores/vercel_postgres.d.ts

@@ -2,7 +2,7 @@ import { type VercelPool, type VercelPoolClient, type VercelPostgresPoolConfig }
 import { VectorStore } from "./base.js";
 import { Embeddings } from "../embeddings/base.js";
 import { Document } from "../document.js";
-type Metadata = Record<string, unknown>;
+type Metadata = Record<string, string | number | Record<"in", string[]>>;
 /**
  * Interface that defines the arguments required to create a
  * `VercelPostgres` instance. It includes Postgres connection options,

package/dist/vectorstores/vercel_postgres.js

@@ -203,14 +203,36 @@ export class VercelPostgres extends VectorStore {
      */
     async similaritySearchVectorWithScore(query, k, filter) {
         const embeddingString = `[${query.join(",")}]`;
-        const _filter = filter ?? "{}";
+        const _filter = filter ?? {};
+        const whereClauses = [];
+        const values = [embeddingString, k];
+        let paramCount = values.length;
+        for (const [key, value] of Object.entries(_filter)) {
+            if (typeof value === "object" && value !== null) {
+                const currentParamCount = paramCount;
+                const placeholders = value.in
+                    .map((_, index) => `$${currentParamCount + index + 1}`)
+                    .join(",");
+                whereClauses.push(`${this.metadataColumnName}->>'${key}' IN (${placeholders})`);
+                values.push(...value.in);
+                paramCount += value.in.length;
+            }
+            else {
+                paramCount += 1;
+                whereClauses.push(`${this.metadataColumnName}->>'${key}' = $${paramCount}`);
+                values.push(value);
+            }
+        }
+        const whereClause = whereClauses.length
+            ? `WHERE ${whereClauses.join(" AND ")}`
+            : "";
         const queryString = `
-      SELECT *, ${this.vectorColumnName} <=> $1 as "_distance"
-      FROM ${this.tableName}
-      WHERE ${this.metadataColumnName} @> $2
-      ORDER BY "_distance" ASC
-      LIMIT $3;`;
-        const documents = (await this.client.query(queryString, [embeddingString, _filter, k])).rows;
+            SELECT *, ${this.vectorColumnName} <=> $1 as "_distance"
+            FROM ${this.tableName}
+            ${whereClause}
+            ORDER BY "_distance" ASC
+            LIMIT $2;`;
+        const documents = (await this.client.query(queryString, values)).rows;
         const results = [];
         for (const doc of documents) {
             if (doc._distance != null && doc[this.contentColumnName] != null) {