开发arXiv论文引擎机器人程序：基于RAG＋LangChain＋Chainlit＋ChromaDB_langchain 读arxiv

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">conda create -n semantic_research_engine pythnotallow=3.10</code></span></span></span>

然后，使用如下命令激活环境：

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">conda activate semantic_research_engine</code></span></span></span>

最后，通过运行以下命令，在激活的环境中安装所有必需的依赖项：

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">pip install -r requirements.txt</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">OPENAI_API_KEY="your_openai_api_key"</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"># rag_test.py 
from langchain_community.document_loaders import ArxivLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Chroma
from langchain_community.embeddings import HuggingFaceEmbeddings

query = "lightweight transformer for language tasks"
arxiv_docs = ArxivLoader(query=query, load_max_docs=3).load()
pdf_data = []
for doc in arxiv_docs:
    text_splitter = RecursiveCharacterTextSplitter(
                    chunk_size=1000,
                    chunk_overlap=100)
    texts = text_splitter.create_documents([doc.page_content])
    pdf_data.append(texts)

embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-l6-v2")
db = Chroma.from_documents(pdf_data[0], embeddings)</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"># rag_test.py
from langchain.chains import RetrievalQA
from langchain_openai import OpenAI
from dotenv import load_dotenv

load_dotenv()
llm = OpenAI(model='gpt-3.5-turbo-instruct', temperature=0)
qa = RetrievalQA.from_chain_type(llm=llm, 
                                 chain_type="stuff", 
                                 retriever=db.as_retriever())

question = "how many and which benchmark datasets and tasks were 
            compared for light weight transformer?"
result = qa({"query": question})</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">LITERAL_API_KEY="your_literal_api_key"</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">You are a helpful assistant. Use provided {<!-- -->{context}} to answer user 
{<!-- -->{question}}. Do not use prior knowledge. 
Answer:</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">#search_engine.py
from literalai import LiteralClient
from dotenv import load_dotenv

load_dotenv()

client = LiteralClient()

# 下面的代码将会使用最新版本，你也可以选择使用特定版本
prompt = client.api.get_prompt(name="test_prompt")
prompt = prompt.to_langchain_chat_prompt_template()
prompt.input_variables = ["context", "question"]</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"># search_engine.py
import chainlit as cl
from langchain_community.document_loaders import ArxivLoader
from langchain_community.vectorstores import Chroma
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.chains import RetrievalQA
from langchain_openai import ChatOpenAI
from dotenv import load_dotenv

load_dotenv()</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"># search_engine.py
# contd.

@cl.on_chat_start
async def retrieve_docs():
    # QUERY PORTION
    arxiv_query = None
    # 等待用户输入一个标题
    while arxiv_query is None:
        arxiv_query = await cl.AskUserMessage(
            content="Please enter a topic to begin!", timeout=15).send()
    query = arxiv_query['output']

    # ARXIV文档部分
    arxiv_docs = ArxivLoader(query=arxiv_query, load_max_docs=3).load()
    # 准备arXiv结果，用于显示
    arxiv_papers = [f"Published: {doc.metadata['Published']} \n "
                    f"Title: {doc.metadata['Title']} \n "
                    f"Authors: {doc.metadata['Authors']} \n "
                    f"Summary: {doc.metadata['Summary'][:50]}... \n---\n"
                    for doc in arxiv_docs]

    await cl.Message(content=f"{arxiv_papers}").send()

    await cl.Message(content=f"Downloading and chunking articles for {query} "
                             f"This operation can take a while!").send()

    #数据库部分
    pdf_data = []
    for doc in arxiv_docs:
        text_splitter = RecursiveCharacterTextSplitter(
                        chunk_size=1000, chunk_overlap=100)
        texts = text_splitter.create_documents([doc.page_content])
        pdf_data.append(texts)
  
    llm = ChatOpenAI(model='gpt-3.5-turbo',
                         temperature=0)
    embeddings = HuggingFaceEmbeddings(
                 model_name="sentence-transformers/all-MiniLM-l6-v2")
    db = Chroma.from_documents(pdf_data[0], embeddings)

    # 链部分
    chain = RetrievalQA.from_chain_type(llm=llm,
                                            chain_type="stuff",
                                            retriever=db.as_retriever(),
                                            chain_type_kwargs={
                                                "verbose": True,
                                                "prompt": prompt
                                            }
                                            )

    # 让用户知道管道已经准备到位
    await cl.Message(content=f"Database creation for `{query}` complete. "
                             f"You can now ask questions!").send()

    cl.user_session.set("chain", chain)
    cl.user_session.set("db", db)</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">@cl.on_message
async def retrieve_docs(message: cl.Message):
    question = message.content
    chain = cl.user_session.get("chain")
    db = cl.user_session.get("db")
    # 为每次调用创建一个新的回调处理程序实例
    cb = client.langchain_callback()
    variables = {"context": db.as_retriever(search_kwargs={"k": 1}), 
                 "query": question}
    database_results = await chain.acall(variables,
                                         callbacks=[cb])
    results = [f"Question: {question} "
               f"\n Answer: {database_results['result']}"]
    await cl.Message(results).send()</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"><!DOCTYPE html>
<html>
  <head>
    <title>Semantic Search Engine</title>
  </head>
 <body>
   <!-- ... -->
   <script src="http://localhost:8000/copilot/index.js"></script>
   <script>
     window.mountChainlitWidget({
       chainlitServer: "http://localhost:8000",
     });
   </script>
 </body></code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">chainlit run search_engine.py -w</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">npx http-server</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">@cl.on_message
async def retrieve_docs(message: cl.Message):
    if cl.context.session.client_type == "copilot":
        # code to be executed only inside the Copilot</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">@cl.on_chat_start
async def retrieve_docs():
    if cl.context.session.client_type == "copilot":
        # 代码同上

        #触发显示arXiv论文结果的弹出窗口
        fn_arxiv = cl.CopilotFunction(name="showArxivResults", 
                   args={"results": "\n".join(arxiv_papers)})
        await fn_arxiv.acall()
        
        # same code as befor</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">@cl.on_message
async def retrieve_docs(message: cl.Message):
    if cl.context.session.client_type == "copilot":
        #代码同上
        
        # 激活相应于数据库查询结果的弹出小窗口
        fn_db = cl.CopilotFunction(name="showDatabaseResults", 
                args={"results": "\n".join(results)})
        await fn_db.acall()
        
        #代码同前</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"><script>
 // 前面的代码
 window.addEventListener("chainlit-call-fn", (e) => {
 const { name, args, callback } = e.detail;
 if (name === "showArxivResults") {
   document.getElementById("arxiv-result-text").innerHTML =
     args.results.replace(/\n/g, "<br>");
   document.getElementById("popup").style.display = "flex";
   if (callback) callback();
 } else if (name === "showDatabaseResults") {
   document.getElementById("database-results-text").innerHTML =
     args.results.replace(/\n/g, "<br>");
   document.getElementById("popup").style.display = "flex";
   if (callback) callback();
 }
 });
</script></code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"><div id="popup" class="popup">
 <span class="close-btn" onclick="hidePopup()">×</span>
 <div class="arxiv-results-wrapper">
   <h1>Arxiv Results</h1>
   <p id="arxiv-result-text">Online results will be displayed here.</p>
 </div>
 <div class="database-results-wrapper">
   <h1>Database Results</h1>
   <p id="database-results-text">Database results will be displayed here.</p>
 </div>
</div></code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs"><style>
 * {
   box-sizing: border-box;
 }
 
 body {
   font-family: sans-serif;
 }
 
 .close-btn {
   position: absolute;
   top: 10px;
   right: 20px;
   font-size: 24px;
   cursor: pointer;
 }
 
 .popup {
   display: none;
   position: fixed;
   top: 50%;
   left: 50%;
   transform: translate(-50%, -50%);
   background-color: white;
   padding: 20px;
   box-shadow: rgba(99, 99, 99, 0.2) 0px 2px 8px 0px;
   width: 40%;
   flex-direction: column;
   gap: 50px;
 }
 
 p {
   color: #00000099;
 }
</style></code></span></span></span>​​​​​

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">chainlit run search.py -h</code></span></span></span>

<span style="color:#333333;"><span style="background-color:#f7f7f7;"><span style="color:#000000;"><code class="language-plain hljs">npx http-server</code></span></span></span>