這邊就是先把我們需要的文件做一次indexing ，把需要的文件轉成向量存入db，可以參考Day9 如何實作，在FastAPI我們就以原先配置DB連線設定，並且轉成API方式可以讓各位自己丟檔案，並且使用pypdf及PyPDFLoader 可以真正的讀取pdf文件！

以下為indexing程式：

先安裝新的兩個套件

pip install python-multipart
pip install pypdf

並且實作一下indexing的API~

@app.post("/indexing")
async def indexing(file: UploadFile = File(...),):

    embeddings = OpenAIEmbeddings()
    vectorstore = PGVector(    
            embeddings=embeddings,
            collection_name="programers",
            connection=CONNECTION_STRING,
            use_jsonb=True)

    temp_file_path = file.filename
    try:
        with open(temp_file_path, "wb") as buffer:
            shutil.copyfileobj(file.file, buffer)

        loader = PyPDFLoader(temp_file_path)

        docs = loader.load()
        os.remove(temp_file_path)
        text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
        splits = text_splitter.split_documents(docs)
        vectorstore.add_documents(docs)
        return {"upload": file.filename}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

最後放個愛因斯坦的wiki來執行結果

基本上DB有我們的數據了~

最後就是把Day12 LLM的串接API方式做最後的完善！把整個鏈都串起來！

以下為增加檢索(Retrieval)完善llm api程式：

@app.post("/chat")
async def chat(message: str):
    try:
        retriever = vectorstore.as_retriever()
        chain = (
            {"context": retriever | _format_docs, "message": RunnablePassthrough()}
            | prompt
            | llm
            | StrOutputParser()
        )
        response = chain.invoke(message)
        return {"response": response}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

最後來問一下問題看一下我們的AI回答的如何

看來有照著我的db去讀取相關資訊出來！大成功！

最後整理一下我的程式碼XD(上面api部份東西抽出來)

import shutil
from fastapi import FastAPI, File, HTTPException, UploadFile
import os
from langchain_community.document_loaders.pdf import PyPDFLoader
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_core.prompts import PromptTemplate
from langchain_postgres import PGVector
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
app = FastAPI()
OPENAI_API_KEY = os.getenv('OPENAI_API_KEY')
CONNECTION_STRING = "postgresql+psycopg://benson:benson@localhost:5432/postgres"

prompt = PromptTemplate.from_template("你是專業工程師 ，根據以下內容{context}，幫我回答以下問題 {message}:\n")

llm = ChatOpenAI(model='gpt-4o-mini')

embeddings = OpenAIEmbeddings()
vectorstore = PGVector(    
        embeddings=embeddings,
        collection_name="programers",
        connection=CONNECTION_STRING,
        use_jsonb=True)

@app.post("/chat")
async def chat(message: str):
    try:
        retriever = vectorstore.as_retriever()
        chain = (
            {"context": retriever | _format_docs, "message": RunnablePassthrough()}
            | prompt
            | llm
            | StrOutputParser()
        )
        response = chain.invoke(message)
        return {"response": response}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
    
@app.post("/indexing")
async def indexing(file: UploadFile = File(...),):

    temp_file_path = file.filename
    try:
        with open(temp_file_path, "wb") as buffer:
            shutil.copyfileobj(file.file, buffer)

        loader = PyPDFLoader(temp_file_path)

        docs = loader.load()
        os.remove(temp_file_path)
        text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
        splits = text_splitter.split_documents(docs)
        vectorstore.add_documents(docs)
        return {"upload": file.filename}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

def _format_docs(docs):
    return "\n\n".join(doc.page_content for doc in docs)

是不是在地端能做一個自己的RAG系統是一件很簡單的事呢！後續會把這套系統配合GCP的方式部署到雲端上，讓各位都可以做使用，後續就會介紹雲端元件，依照後續章節也會迭代調整程式碼！

各位要好好看完！我後續在某一天會把這個git實作開源出來XD