Zep 开源项目

Zep 是一个为 AI 助手应用提供的长期记忆服务。 使用 Zep，你可以让 AI 助手具备回忆过去对话的能力，无论这些对话多么久远， 同时还能减少幻觉、降低延迟和成本。

对 Zep Cloud 感兴趣？请参阅 Zep Cloud 安装指南

注意： ZepVectorStore 与 Documents 一起工作，用于作为 Retriever 使用。 它与 Zep 的 ZepMemory 类提供不同的功能，后者用于持久化、丰富和搜索用户的聊天历史。

为什么选择 Zep 的 VectorStore？🤖🚀

Zep 使用部署在 Zep 服务器本地的低延迟模型，自动对添加到 Zep Vector Store 的文档进行嵌入。 Zep 的 TypeScript/JavaScript 客户端可以在非 Node 的边缘环境中使用。 这两者结合 Zep 的聊天记忆功能，使 Zep 成为构建对延迟和性能要求较高的对话式 LLM 应用的理想选择。

支持的搜索类型

Zep 支持相似性搜索和最大边界相关性（MMR）搜索。MMR 搜索在检索增强生成（RAG）应用中特别有用，因为它会对结果重新排序，以确保返回的文档具有多样性。

安装

请参考 Zep 开源快速入门指南，了解如何安装和开始使用 Zep。

使用方法

你需要提供 Zep API 的 URL，以及可选的 API 密钥，以使用 Zep VectorStore。更多信息请参见 Zep 文档。

在下面的示例中，我们使用了 Zep 的自动嵌入功能，它会使用 Zep 服务器上的低延迟嵌入模型自动对文档进行嵌入。由于 LangChain 需要传入一个 Embeddings 实例，我们传入了 FakeEmbeddings。

注意： 如果你传入的 Embeddings 实例不是 FakeEmbeddings，则该类将用于对文档进行嵌入。 你还必须将文档集合的 isAutoEmbedded 设置为 false。请参见下面的 OpenAIEmbeddings 示例。

示例：从文档创建 ZepVectorStore 并进行查询

:::提示 请参阅安装集成包的一般说明部分。 :::

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

import { ZepVectorStore } from "@langchain/community/vectorstores/zep";
import { FakeEmbeddings } from "@langchain/core/utils/testing";
import { TextLoader } from "langchain/document_loaders/fs/text";
import { randomUUID } from "crypto";

const loader = new TextLoader("src/document_loaders/example_data/example.txt");
const docs = await loader.load();
export const run = async () => {
  const collectionName = `collection${randomUUID().split("-")[0]}`;

  const zepConfig = {
    apiUrl: "http://localhost:8000", // this should be the URL of your Zep implementation
    collectionName,
    embeddingDimensions: 1536, // this much match the width of the embeddings you're using
    isAutoEmbedded: true, // If true, the vector store will automatically embed documents when they are added
  };

  const embeddings = new FakeEmbeddings();

  const vectorStore = await ZepVectorStore.fromDocuments(
    docs,
    embeddings,
    zepConfig
  );

  // Wait for the documents to be embedded
  // eslint-disable-next-line no-constant-condition
  while (true) {
    const c = await vectorStore.client.document.getCollection(collectionName);
    console.log(
      `Embedding status: ${c.document_embedded_count}/${c.document_count} documents embedded`
    );
    // eslint-disable-next-line no-promise-executor-return
    await new Promise((resolve) => setTimeout(resolve, 1000));
    if (c.status === "ready") {
      break;
    }
  }

  const results = await vectorStore.similaritySearchWithScore("bar", 3);

  console.log("Similarity Results:");
  console.log(JSON.stringify(results));

  const results2 = await vectorStore.maxMarginalRelevanceSearch("bar", {
    k: 3,
  });

  console.log("MMR Results:");
  console.log(JSON.stringify(results2));
};

API Reference:

• ZepVectorStore from @langchain/community/vectorstores/zep
• FakeEmbeddings from @langchain/core/utils/testing
• TextLoader from langchain/document_loaders/fs/text

示例：使用元数据过滤器查询 ZepVectorStore

import { ZepVectorStore } from "@langchain/community/vectorstores/zep";
import { FakeEmbeddings } from "@langchain/core/utils/testing";
import { randomUUID } from "crypto";
import { Document } from "@langchain/core/documents";

const docs = [
  new Document({
    metadata: { album: "Led Zeppelin IV", year: 1971 },
    pageContent:
      "Stairway to Heaven is one of the most iconic songs by Led Zeppelin.",
  }),
  new Document({
    metadata: { album: "Led Zeppelin I", year: 1969 },
    pageContent:
      "Dazed and Confused was a standout track on Led Zeppelin's debut album.",
  }),
  new Document({
    metadata: { album: "Physical Graffiti", year: 1975 },
    pageContent:
      "Kashmir, from Physical Graffiti, showcases Led Zeppelin's unique blend of rock and world music.",
  }),
  new Document({
    metadata: { album: "Houses of the Holy", year: 1973 },
    pageContent:
      "The Rain Song is a beautiful, melancholic piece from Houses of the Holy.",
  }),
  new Document({
    metadata: { band: "Black Sabbath", album: "Paranoid", year: 1970 },
    pageContent:
      "Paranoid is Black Sabbath's second studio album and includes some of their most notable songs.",
  }),
  new Document({
    metadata: {
      band: "Iron Maiden",
      album: "The Number of the Beast",
      year: 1982,
    },
    pageContent:
      "The Number of the Beast is often considered Iron Maiden's best album.",
  }),
  new Document({
    metadata: { band: "Metallica", album: "Master of Puppets", year: 1986 },
    pageContent:
      "Master of Puppets is widely regarded as Metallica's finest work.",
  }),
  new Document({
    metadata: { band: "Megadeth", album: "Rust in Peace", year: 1990 },
    pageContent:
      "Rust in Peace is Megadeth's fourth studio album and features intricate guitar work.",
  }),
];

export const run = async () => {
  const collectionName = `collection${randomUUID().split("-")[0]}`;

  const zepConfig = {
    apiUrl: "http://localhost:8000", // this should be the URL of your Zep implementation
    collectionName,
    embeddingDimensions: 1536, // this much match the width of the embeddings you're using
    isAutoEmbedded: true, // If true, the vector store will automatically embed documents when they are added
  };

  const embeddings = new FakeEmbeddings();

  const vectorStore = await ZepVectorStore.fromDocuments(
    docs,
    embeddings,
    zepConfig
  );

  // Wait for the documents to be embedded
  // eslint-disable-next-line no-constant-condition
  while (true) {
    const c = await vectorStore.client.document.getCollection(collectionName);
    console.log(
      `Embedding status: ${c.document_embedded_count}/${c.document_count} documents embedded`
    );
    // eslint-disable-next-line no-promise-executor-return
    await new Promise((resolve) => setTimeout(resolve, 1000));
    if (c.status === "ready") {
      break;
    }
  }

  vectorStore
    .similaritySearchWithScore("sad music", 3, {
      where: { jsonpath: "$[*] ? (@.year == 1973)" }, // We should see a single result: The Rain Song
    })
    .then((results) => {
      console.log(`\n\nSimilarity Results:\n${JSON.stringify(results)}`);
    })
    .catch((e) => {
      if (e.name === "NotFoundError") {
        console.log("No results found");
      } else {
        throw e;
      }
    });

  // We're not filtering here, but rather demonstrating MMR at work.
  // We could also add a filter to the MMR search, as we did with the similarity search above.
  vectorStore
    .maxMarginalRelevanceSearch("sad music", {
      k: 3,
    })
    .then((results) => {
      console.log(`\n\nMMR Results:\n${JSON.stringify(results)}`);
    })
    .catch((e) => {
      if (e.name === "NotFoundError") {
        console.log("No results found");
      } else {
        throw e;
      }
    });
};

API Reference:

• ZepVectorStore from @langchain/community/vectorstores/zep
• FakeEmbeddings from @langchain/core/utils/testing
• Document from @langchain/core/documents

示例：使用 LangChain 的嵌入类，例如 OpenAIEmbeddings

import { ZepVectorStore } from "@langchain/community/vectorstores/zep";
import { OpenAIEmbeddings } from "@langchain/openai";
import { TextLoader } from "langchain/document_loaders/fs/text";
import { randomUUID } from "crypto";

const loader = new TextLoader("src/document_loaders/example_data/example.txt");
const docs = await loader.load();
export const run = async () => {
  const collectionName = `collection${randomUUID().split("-")[0]}`;

  const zepConfig = {
    apiUrl: "http://localhost:8000", // this should be the URL of your Zep implementation
    collectionName,
    embeddingDimensions: 1536, // this much match the width of the embeddings you're using
    isAutoEmbedded: false, // set to false to disable auto-embedding
  };

  const embeddings = new OpenAIEmbeddings();

  const vectorStore = await ZepVectorStore.fromDocuments(
    docs,
    embeddings,
    zepConfig
  );

  const results = await vectorStore.similaritySearchWithScore("bar", 3);

  console.log("Similarity Results:");
  console.log(JSON.stringify(results));

  const results2 = await vectorStore.maxMarginalRelevanceSearch("bar", {
    k: 3,
  });

  console.log("MMR Results:");
  console.log(JSON.stringify(results2));
};

API Reference:

• ZepVectorStore from @langchain/community/vectorstores/zep
• OpenAIEmbeddings from @langchain/openai
• TextLoader from langchain/document_loaders/fs/text

相关内容

• 向量存储 概念指南
• 向量存储 操作指南

Related

• Vector store conceptual guide
• Vector store how-to guides