Zep Cloud Retriever

Zep 是 AI 助手应用的长期记忆服务。借助 Zep，您可以为 AI 助手提供回忆过去对话的能力，无论对话多么久远，同时还可以减少幻觉、延迟和成本。

此示例展示了如何在检索链中使用 Zep Retriever 从 Zep Open Source 内存存储中检索文档。

安装

注册 Zep Cloud 并创建一个项目。

按照 Zep Cloud Typescript SDK 安装指南 进行安装并开始使用 Zep。

您需要您的 Zep Cloud 项目 API 密钥才能使用 ZepCloudRetriever。有关更多信息，请参阅 Zep Cloud 文档。

设置

提示

有关安装集成包的一般说明，请参阅此部分。

npm

npm i @getzep/zep-cloud @langchain/community @langchain/core

Yarn

yarn add @getzep/zep-cloud @langchain/community @langchain/core

pnpm

pnpm add @getzep/zep-cloud @langchain/community @langchain/core

用法

import { ZepCloudRetriever } from "@langchain/community/retrievers/zep_cloud";
import { randomUUID } from "crypto";
import { ZepClient } from "@getzep/zep-cloud";
import { Message } from "@getzep/zep-cloud/api";

function sleep(ms: number) {
  // eslint-disable-next-line no-promise-executor-return
  return new Promise((resolve) => setTimeout(resolve, ms));
}

const zepConfig = {
  // Your Zep Cloud Project API key https://help.getzep.com/projects
  apiKey: "<Zep Api Key>",
  sessionId: `session_${randomUUID()}`,
};

console.log(`Zep Config: ${JSON.stringify(zepConfig)}`);

// Generate chat messages about traveling to France
const chatMessages = [
  {
    role: "AI",
    message: "Bonjour! How can I assist you with your travel plans today?",
  },
  { role: "User", message: "I'm planning a trip to France." },
  {
    role: "AI",
    message: "That sounds exciting! What cities are you planning to visit?",
  },
  { role: "User", message: "I'm thinking of visiting Paris and Nice." },
  {
    role: "AI",
    message: "Great choices! Are you interested in any specific activities?",
  },
  { role: "User", message: "I would love to visit some vineyards." },
  {
    role: "AI",
    message:
      "France has some of the best vineyards in the world. I can help you find some.",
  },
  { role: "User", message: "That would be great!" },
  { role: "AI", message: "Do you prefer red or white wine?" },
  { role: "User", message: "I prefer red wine." },
  {
    role: "AI",
    message:
      "Perfect! I'll find some vineyards that are known for their red wines.",
  },
  { role: "User", message: "Thank you, that would be very helpful." },
  {
    role: "AI",
    message:
      "You're welcome! I'll also look up some French wine etiquette for you.",
  },
  {
    role: "User",
    message: "That sounds great. I can't wait to start my trip!",
  },
  {
    role: "AI",
    message:
      "I'm sure you'll have a fantastic time. Do you have any other questions about your trip?",
  },
  { role: "User", message: "Not at the moment, thank you for your help!" },
];

const zepClient = new ZepClient({
  apiKey: zepConfig.apiKey,
});

// Add chat messages to memory
for (const chatMessage of chatMessages) {
  let m: Message;
  if (chatMessage.role === "AI") {
    m = { role: "ai", roleType: "assistant", content: chatMessage.message };
  } else {
    m = { role: "human", roleType: "user", content: chatMessage.message };
  }

  await zepClient.memory.add(zepConfig.sessionId, { messages: [m] });
}

// Wait for messages to be summarized, enriched, embedded and indexed.
await sleep(10000);

// Simple similarity search
const query = "Can I drive red cars in France?";
const retriever = new ZepCloudRetriever({ ...zepConfig, topK: 3 });
const docs = await retriever.invoke(query);
console.log("Simple similarity search");
console.log(JSON.stringify(docs, null, 2));

// mmr reranking search
const mmrRetriever = new ZepCloudRetriever({
  ...zepConfig,
  topK: 3,
  searchType: "mmr",
  mmrLambda: 0.5,
});
const mmrDocs = await mmrRetriever.invoke(query);
console.log("MMR reranking search");
console.log(JSON.stringify(mmrDocs, null, 2));

// summary search with mmr reranking
const mmrSummaryRetriever = new ZepCloudRetriever({
  ...zepConfig,
  topK: 3,
  searchScope: "summary",
  searchType: "mmr",
  mmrLambda: 0.5,
});
const mmrSummaryDocs = await mmrSummaryRetriever.invoke(query);
console.log("Summary search with MMR reranking");
console.log(JSON.stringify(mmrSummaryDocs, null, 2));

// Filtered search
const filteredRetriever = new ZepCloudRetriever({
  ...zepConfig,
  topK: 3,
  filter: {
    where: { jsonpath: '$[*] ? (@.foo == "bar")' },
  },
});
const filteredDocs = await filteredRetriever.invoke(query);
console.log("Filtered search");
console.log(JSON.stringify(filteredDocs, null, 2));

API 参考

• ZepCloudRetriever 来自 @langchain/community/retrievers/zep_cloud

相关

• 检索器 概念指南
• 检索器 操作指南