LangChain + OpenAI 进行文档分析

发表于 更新于 分类于 Waline：

最近在玩Agent，用LangChain写了一个简单的DEMO，可以实现根据输入文档和之前的聊天记录来进行问答，之所以记录一下，实在是因为Langchain的文档非常差，这段代码耗时三天才成功运行。

我的代码是运行在自己的NextJS项目中，所以最后会有些返回请求体的代码

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
import { OpenAI, OpenAIEmbeddings, ChatOpenAI } from "@langchain/openai";
// import { Ollama, OllamaEmbeddings } from "@langchain/ollama";
import { ConversationalRetrievalQAChain } from "langchain/chains";
import { HNSWLib } from "@langchain/community/vectorstores/hnswlib";
import { RecursiveCharacterTextSplitter } from "@langchain/textsplitters";
import { BufferMemory } from "langchain/memory";
export const createEmbeddings = async (req: NextRequest) => {
  const { id } = getRannieServerSession(req);
  const form = await req.formData();
  const vectorStoreIndex = form.get("vectorStoreIndex") as string;
  const vectorStoreDirectory = `${process.cwd()}/temp/${id}-${vectorStoreIndex}`;
  vectorStoreDirectoryProcessing.add(vectorStoreDirectory);
  /* Initialize the LLM to use to answer the question */
  // const model = new ChatOpenAI({ model: 'gpt-4o', streaming: true, apiKey: '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' });
  /* Load in the file we want to do question answering over */
  /* Split the text into chunks */
  const textSplitter = new RecursiveCharacterTextSplitter({ chunkSize: 1000 });
  /* Create the vectorstore */
  const embedding = new OpenAIEmbeddings({
    model: "text-embedding-3-large",
    apiKey: getApiKey(),
  });
  const file = form.get("file") as File;
  const text = await readFileContent(file);
  const docs = await textSplitter.createDocuments([text]);
  let vectorStore: HNSWLib;
  if (fs.existsSync(vectorStoreDirectory)) {
    vectorStore = await HNSWLib.load(vectorStoreDirectory, embedding);
    await vectorStore.addDocuments(docs);
  } else {
    createDirectoryRecursively(vectorStoreDirectory);
    vectorStore = await HNSWLib.fromDocuments(docs, embedding);
  }
  await vectorStore.save(vectorStoreDirectory);
  vectorStoreDirectoryProcessing.delete(vectorStoreDirectory);

  return true;
};

export async function qaDocument(req: NextRequest) {
  const requestBody = await req.clone().json();
  const { id } = getRannieServerSession(req);
  const { vectorStoreIndex } = requestBody ?? {};
  const vectorStoreDirectory = `${process.cwd()}/temp/${id}-${vectorStoreIndex}`;

  if (
    !fs.existsSync(vectorStoreDirectory) ||
    vectorStoreDirectoryProcessing.has(vectorStoreDirectory)
  ) {
    return NextResponse.json(getError(30004), {
      status: 400,
    });
  }

  /* Initialize the LLM to use to answer the question */
  const model = new ChatOpenAI({
    model: "gpt-4o",
    streaming: true,
    streamUsage: true,
    apiKey: getApiKey(),
  });
  const embedding = new OpenAIEmbeddings({
    model: "text-embedding-3-large",
    apiKey: getApiKey(),
  });
  const vectorStore = await HNSWLib.load(vectorStoreDirectory, embedding);

  const bufferMemory = new BufferMemory({
    memoryKey: "chat_history", // Must be set to "chat_history"
  });
  let question = "";
  for (let i = 0; i < requestBody.messages.length; i++) {
    const message = requestBody.messages[i];
    if (message.role === "system") {
      bufferMemory.chatHistory.addMessage(new SystemMessage(message));
    } else if (message.role === "user") {
      if (i === requestBody.messages.length - 1) {
        question = message.content as string;
      } else {
        bufferMemory.chatHistory.addMessage(new HumanMessage(message));
      }
    } else if (message.role === "assistant") {
      bufferMemory.chatHistory.addMessage(
        new AIMessage({ content: message.content || "" }),
      );
    }
  }
  /* Create the chain */
  const chain = ConversationalRetrievalQAChain.fromLLM(
    model,
    vectorStore.asRetriever(),
    {
      memory: bufferMemory,
      qaChainOptions: {
        type: "stuff",
      },
    },
  );
  const body = new ReadableStream({
    start(controller) {
      chain
        .invoke(
          { question },
          {
            callbacks: [
              // choices[0].delta.content
              {
                handleLLMNewToken: (token) =>
                  controller.enqueue(
                    `data: ${JSON.stringify({
                      choices: [{ delta: { content: token } }],
                    })}\n\n`,
                  ),
                handleLLMEnd(output) {
                },
              },
            ],
          },
        )
        .then((res) => {
          controller.enqueue("data: [DONE]");
        })
        .finally(() => {
          controller.close();
        });
    },
  });
  const newHeaders = new Headers();
  newHeaders.set("Content-Type", "text/event-stream");
  newHeaders.set("Cache-Control", "no-cache");
  newHeaders.set("Connection", "keep-alive");
  return new Response(body, {
    headers: newHeaders,
  });
}

上述代码分为两部分，第一部分是根据file创建Hnswlib向量，创建完成后保存下来，后续如果需要使用，直接从本地加载