Zep 是一种用于 AI 助手应用程序的长期记忆服务。通过 Zep,您可以让 AI 助手回忆过去的对话,无论它们有多遥远,同时还能减少幻觉、延迟和成本。
对 Zep 云感兴趣?请参阅 Zep 云安装指南此示例展示了如何使用 Zep 检索器在检索链中从 Zep 开源内存存储中检索文档。
安装
请遵循 Zep 开源快速入门指南 来安装和开始使用 Zep。设置
有关安装 LangChain 软件包的一般说明,请参阅此部分。
npm
复制
向 AI 提问
npm i @getzep/zep-js @langchain/community @langchain/core
用法
复制
向 AI 提问
import { ZepRetriever } from "@langchain/community/retrievers/zep";
import { ZepMemory } from "@langchain/community/memory/zep";
import { Memory as MemoryModel, Message } from "@getzep/zep-js";
import { randomUUID } from "crypto";
function sleep(ms: number) {
// eslint-disable-next-line no-promise-executor-return
return new Promise((resolve) => setTimeout(resolve, ms));
}
export const run = async () => {
const zepConfig = {
url: process.env.ZEP_URL || "https://:8000",
sessionId: `session_${randomUUID()}`,
};
console.log(`Zep Config: ${JSON.stringify(zepConfig)}`);
const memory = new ZepMemory({
baseURL: zepConfig.url,
sessionId: zepConfig.sessionId,
});
// 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 = await memory.zepClientPromise;
if (!zepClient) {
throw new Error("ZepClient is not initialized");
}
// Add chat messages to memory
for (const chatMessage of chatMessages) {
let m: MemoryModel;
if (chatMessage.role === "AI") {
m = new MemoryModel({
messages: [new Message({ role: "ai", content: chatMessage.message })],
});
} else {
m = new MemoryModel({
messages: [
new Message({ role: "human", content: chatMessage.message }),
],
});
}
await zepClient.memory.addMemory(zepConfig.sessionId, 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 ZepRetriever({ ...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 ZepRetriever({
...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 ZepRetriever({
...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 ZepRetriever({
...zepConfig,
topK: 3,
filter: {
where: { jsonpath: '$.system.entities[*] ? (@.Label == "GPE")' },
},
});
const filteredDocs = await filteredRetriever.invoke(query);
console.log("Filtered search");
console.log(JSON.stringify(filteredDocs, null, 2));
};
相关
以编程方式连接这些文档到 Claude、VSCode 等,通过 MCP 获取实时答案。