使用 wechaty langchain 部署私有 chatgpt

WeChaty 是一个基于 Node.js 的开源微信机器人框架，而 LangChain 是一个用于部署私有化 GPT 模型的工具。通过结合 WeChaty 和 LangChain，你可以创建一个私有化的 GPT 机器人，使其在微信平台上运行。

Setup：

我们使用 wechaty-puppet-wechat4u

package.json:
"wechaty": "^1.20.2",
"wechaty-puppet-wechat4u": "1.14.1"
"langchain": "^0.0.102",
"@pinecone-database/pinecone": "^0.1.6",
"pdf-parse": "^1.1.1", // 篇幅原因这里只演示 pdf

import { WechatyBuilder} from 'wechaty'

const wechaty = WechatyBuilder.build({
  name: 'wechaty-chatgpt',
  puppet: 'wechaty-puppet-wechat4u',
  puppetOptions: {
    uos: true,
  },
});

设置 pinecone ,openai

PROMPTLAYER_API_KEY=pl_... # PROMPTLAYER 是一个用于记录 api 调用时 prompt 与 response 的工具
PINECONE_API_KEY=89e...
PINECONE_ENVIRONMENT=us-west4-gcp-free
PINECONE_INDEX=...

以下代码为当接收到支持的文件对文件进行向量化成功后返回提示

   wechaty.on('message', async message => {
    const contact = message.talker();
    currentAdminUser = contact.payload.alias === process.env.ADMIN
    const receiver = message.listener();
    let content = message.text().trim();
    const room = message.room();
    const target = room || contact;
    const isText = message.type() === wechaty.Message.Type.Text;
    const isAudio = message.type() === wechaty.Message.Type.Audio;
    const isFile = message.type() === wechaty.Message.Type.Attachment;

    if (isFile) {
      const filebox = await message.toFileBox()
      if (supportFileType(filebox.mediaType)) {
        await saveFile(filebox)
        await loadDocuments()
        await send(room || contact, `${filebox.name} Embeddings 成功`)
        return
      }
    }
   })  

langchain 相关代码

  import { PineconeClient } from "@pinecone-database/pinecone";
  import dotenv from 'dotenv';
  import { VectorDBQAChain } from "langchain/chains";
  import { DirectoryLoader } from "langchain/document_loaders";
  import { DocxLoader } from "langchain/document_loaders/fs/docx";
  import { PDFLoader } from "langchain/document_loaders/fs/pdf";
  import { TextLoader } from "langchain/document_loaders/fs/text";
  import { OpenAIEmbeddings } from "langchain/embeddings/openai";
  import { PromptLayerOpenAI } from "langchain/llms/openai";
  import { RecursiveCharacterTextSplitter } from "langchain/text_splitter";
  import { PineconeStore } from "langchain/vectorstores";

  dotenv.config();
  const client = new PineconeClient();

  await client.init({
      apiKey: process.env.PINECONE_API_KEY,
      environment: process.env.PINECONE_ENVIRONMENT,
  });

  const pineconeIndex = client.Index(process.env.PINECONE_INDEX);


  async function loadDocuments(directory = 'resource') {
      console.log('loadDocuments...')
      const loader = new DirectoryLoader(directory,
          {
              ".pdf": (path) => new PDFLoader(path),
              ".txt": (path) => new TextLoader(path),
              ".doc": (path) => new DocxLoader(path),
              ".docx": (path) => new DocxLoader(path),
          });
      // 将数据转成 document 对象，每个文件会作为一个 document
      const rawDocuments = await loader.load();
      console.log(`documents: ${rawDocuments.length}`);

      // 初始化加载器
      const textSplitter = new RecursiveCharacterTextSplitter({ chunkSize: 500 });
      // 切割加载的 document
      const splitDocs = await textSplitter.splitDocuments(rawDocuments);

      // 持久化数据
      // const docsearch = await Chroma.fromDocuments(splitDocs, embeddings, { collectionName: "private_doc" });
      // docsearch.persist();


      await PineconeStore.fromDocuments(splitDocs, new OpenAIEmbeddings(), {
          pineconeIndex,
      });
      console.log(`send to PineconeStore`);

  }


  async function askDocument(question) {
      const llm = new PromptLayerOpenAI({ plTags: ["langchain-requests", "chatbot"] })
      // 初始化 openai 的 embeddings 对象

      // 加载数据
      const vectorStore = await PineconeStore.fromExistingIndex(
          new OpenAIEmbeddings(),
          { pineconeIndex }
      );

      /* Search the vector DB independently with meta filters */
      const chain = VectorDBQAChain.fromLLM(llm, vectorStore, {
          k: 1,
          returnSourceDocuments: true,
      });
      const response = await chain.call({ query: question });
      console.log(response);

      // const response = await vectorStore.similaritySearch(question, 1);
      // console.log(response);

      return response.text
  }

  function supportFileType(mediaType) {
      const types = ['doc', 'docx', , 'pdf', 'text']
      return types.filter(e => mediaType.includes(e)).length > 0
  }


  export { askDocument, loadDocuments, supportFileType };