Neo4j 向量索引

Neo4j 是一个开源图数据库，集成了对向量相似性搜索的支持。它支持

• 近似最近邻搜索
• 欧几里得相似性和余弦相似性
• 结合向量和关键词搜索的混合搜索

安装设置

要使用 Neo4j 向量索引，您需要安装 neo4j-driver 包。

npm

npm install neo4j-driver

Yarn

yarn add neo4j-driver

pnpm

pnpm add neo4j-driver

提示

有关安装集成包的通用说明，请参阅此部分。

npm

npm install @langchain/openai @langchain/community @langchain/core

Yarn

yarn add @langchain/openai @langchain/community @langchain/core

pnpm

pnpm add @langchain/openai @langchain/community @langchain/core

使用 docker-compose 设置 Neo4j 自托管实例

Neo4j 提供了一个预构建的 Docker 镜像，可以用于快速设置自托管的 Neo4j 数据库实例。创建一个名为 docker-compose.yml 的文件，内容如下：

export default {services:{database:{image:'neo4j',ports:['7687:7687','7474:7474'],environment:['NEO4J_AUTH=neo4j/pleaseletmein']}}};

API 参考

然后在同一目录下，运行 docker compose up 启动容器。

您可以在 Neo4j 官网上找到更多关于如何设置 Neo4j 的信息。

用法

以下是使用 Neo4jVectorStore 的一个完整示例：

import { OpenAIEmbeddings } from "@langchain/openai";
import { Neo4jVectorStore } from "@langchain/community/vectorstores/neo4j_vector";

// Configuration object for Neo4j connection and other related settings
const config = {
  url: "bolt://#:7687", // URL for the Neo4j instance
  username: "neo4j", // Username for Neo4j authentication
  password: "pleaseletmein", // Password for Neo4j authentication
  indexName: "vector", // Name of the vector index
  keywordIndexName: "keyword", // Name of the keyword index if using hybrid search
  searchType: "vector" as const, // Type of search (e.g., vector, hybrid)
  nodeLabel: "Chunk", // Label for the nodes in the graph
  textNodeProperty: "text", // Property of the node containing text
  embeddingNodeProperty: "embedding", // Property of the node containing embedding
};

const documents = [
  { pageContent: "what's this", metadata: { a: 2 } },
  { pageContent: "Cat drinks milk", metadata: { a: 1 } },
];

const neo4jVectorIndex = await Neo4jVectorStore.fromDocuments(
  documents,
  new OpenAIEmbeddings(),
  config
);

const results = await neo4jVectorIndex.similaritySearch("water", 1);

console.log(results);

/*
  [ Document { pageContent: 'Cat drinks milk', metadata: { a: 1 } } ]
*/

await neo4jVectorIndex.close();

API 参考

• OpenAIEmbeddings 来自 @langchain/openai
• Neo4jVectorStore 来自 @langchain/community/vectorstores/neo4j_vector

使用 retrievalQuery 参数自定义响应

import { OpenAIEmbeddings } from "@langchain/openai";
import { Neo4jVectorStore } from "@langchain/community/vectorstores/neo4j_vector";

/*
 * The retrievalQuery is a customizable Cypher query fragment used in the Neo4jVectorStore class to define how
 * search results should be retrieved and presented from the Neo4j database. It allows developers to specify
 * the format and structure of the data returned after a similarity search.
 * Mandatory columns for `retrievalQuery`:
 *
 * 1. text:
 *    - Description: Represents the textual content of the node.
 *    - Type: String
 *
 * 2. score:
 *    - Description: Represents the similarity score of the node in relation to the search query. A
 *      higher score indicates a closer match.
 *    - Type: Float (ranging between 0 and 1, where 1 is a perfect match)
 *
 * 3. metadata:
 *    - Description: Contains additional properties and information about the node. This can include
 *      any other attributes of the node that might be relevant to the application.
 *    - Type: Object (key-value pairs)
 *    - Example: { "id": "12345", "category": "Books", "author": "John Doe" }
 *
 * Note: While you can customize the `retrievalQuery` to fetch additional columns or perform
 * transformations, never omit the mandatory columns. The names of these columns (`text`, `score`,
 * and `metadata`) should remain consistent. Renaming them might lead to errors or unexpected behavior.
 */

// Configuration object for Neo4j connection and other related settings
const config = {
  url: "bolt://#:7687", // URL for the Neo4j instance
  username: "neo4j", // Username for Neo4j authentication
  password: "pleaseletmein", // Password for Neo4j authentication
  retrievalQuery: `
    RETURN node.text AS text, score, {a: node.a * 2} AS metadata
  `,
};

const documents = [
  { pageContent: "what's this", metadata: { a: 2 } },
  { pageContent: "Cat drinks milk", metadata: { a: 1 } },
];

const neo4jVectorIndex = await Neo4jVectorStore.fromDocuments(
  documents,
  new OpenAIEmbeddings(),
  config
);

const results = await neo4jVectorIndex.similaritySearch("water", 1);

console.log(results);

/*
  [ Document { pageContent: 'Cat drinks milk', metadata: { a: 2 } } ]
*/

await neo4jVectorIndex.close();

API 参考

• OpenAIEmbeddings 来自 @langchain/openai
• Neo4jVectorStore 来自 @langchain/community/vectorstores/neo4j_vector

从现有图实例化 Neo4jVectorStore

import { OpenAIEmbeddings } from "@langchain/openai";
import { Neo4jVectorStore } from "@langchain/community/vectorstores/neo4j_vector";

/**
 * `fromExistingGraph` Method:
 *
 * Description:
 * This method initializes a `Neo4jVectorStore` instance using an existing graph in the Neo4j database.
 * It's designed to work with nodes that already have textual properties but might not have embeddings.
 * The method will compute and store embeddings for nodes that lack them.
 *
 * Note:
 * This method is particularly useful when you have a pre-existing graph with textual data and you want
 * to enhance it with vector embeddings for similarity searches without altering the original data structure.
 */

// Configuration object for Neo4j connection and other related settings
const config = {
  url: "bolt://#:7687", // URL for the Neo4j instance
  username: "neo4j", // Username for Neo4j authentication
  password: "pleaseletmein", // Password for Neo4j authentication
  indexName: "wikipedia",
  nodeLabel: "Wikipedia",
  textNodeProperties: ["title", "description"],
  embeddingNodeProperty: "embedding",
  searchType: "hybrid" as const,
};

// You should have a populated Neo4j database to use this method
const neo4jVectorIndex = await Neo4jVectorStore.fromExistingGraph(
  new OpenAIEmbeddings(),
  config
);

await neo4jVectorIndex.close();

API 参考

• OpenAIEmbeddings 来自 @langchain/openai
• Neo4jVectorStore 来自 @langchain/community/vectorstores/neo4j_vector

元数据过滤

import { OpenAIEmbeddings } from "@langchain/openai";
import { Neo4jVectorStore } from "@langchain/community/vectorstores/neo4j_vector";

/**
 * `similaritySearch` Method with Metadata Filtering:
 *
 * Description:
 * This method facilitates advanced similarity searches within a Neo4j vector index, leveraging both text embeddings and metadata attributes.
 * The third parameter, `filter`, allows for the specification of metadata-based conditions that pre-filter the nodes before performing the similarity search.
 * This approach enhances the search precision by allowing users to query based on complex metadata criteria alongside textual similarity.
 * Metadata filtering also support the following operators:
 *
 *  $eq: Equal
 *  $ne: Not Equal
 *  $lt: Less than
 *  $lte: Less than or equal
 *  $gt: Greater than
 *  $gte: Greater than or equal
 *  $in: In a list of values
 *  $nin: Not in a list of values
 *  $between: Between two values
 *  $like: Text contains value
 *  $ilike: lowered text contains value
 *
 * The filter supports a range of query operations such as equality checks, range queries, and compound conditions (using logical operators like $and, $or).
 * This makes it highly adaptable to varied use cases requiring detailed and specific retrieval of documents based on both content and contextual information.
 *
 * Note:
 * Effective use of this method requires a well-structured Neo4j database where nodes are enriched with both text and metadata properties.
 * The method is particularly useful in scenarios where the integration of text analysis with detailed metadata querying is crucial, such as in content recommendation systems, detailed archival searches, or any application where contextual relevance is key.
 */

// Configuration object for Neo4j connection and other related settings
const config = {
  url: "bolt://#:7687", // URL for the Neo4j instance
  username: "neo4j", // Username for Neo4j authentication
  password: "pleaseletmein", // Password for Neo4j authentication
  indexName: "vector", // Name of the vector index
  keywordIndexName: "keyword", // Name of the keyword index if using hybrid search
  searchType: "vector" as const, // Type of search (e.g., vector, hybrid)
  nodeLabel: "Chunk", // Label for the nodes in the graph
  textNodeProperty: "text", // Property of the node containing text
  embeddingNodeProperty: "embedding", // Property of the node containing embedding
};

const documents = [
  { pageContent: "what's this", metadata: { a: 2 } },
  { pageContent: "Cat drinks milk", metadata: { a: 1 } },
];

const neo4jVectorIndex = await Neo4jVectorStore.fromDocuments(
  documents,
  new OpenAIEmbeddings(),
  config
);

const filter = { a: { $eq: 1 } };
const results = await neo4jVectorIndex.similaritySearch("water", 1, { filter });

console.log(results);

/*
  [ Document { pageContent: 'Cat drinks milk', metadata: { a: 1 } } ]
*/

await neo4jVectorIndex.close();

API 参考

• OpenAIEmbeddings 来自 @langchain/openai
• Neo4jVectorStore 来自 @langchain/community/vectorstores/neo4j_vector

免责声明 ⚠️

安全提示：请确保数据库连接使用的凭据具有狭窄的范围，仅包含必要的权限。否则可能会导致数据损坏或丢失，因为调用代码可能会尝试执行导致删除、数据突变（如果被适当提示）或读取敏感数据（如果此类数据存在于数据库中）的命令。防止此类负面结果的最佳方法是（在适当的情况下）限制授予与此工具一起使用的凭据的权限。例如，为数据库创建只读用户是确保调用代码无法突变或删除数据的好方法。有关更多信息，请参阅安全页面。

相关内容

• 向量数据库概念指南
• 向量数据库操作指南