rag_system.py

import os
from dotenv import load_dotenv
from langchain_openai import OpenAIEmbeddings
from langchain_chroma import Chroma  # 이 줄을 수정
from langchain.chains import ConversationalRetrievalChain
from langchain_openai import ChatOpenAI
from langchain.docstore.document import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
import pdfplumber
from concurrent.futures import ThreadPoolExecutor

# Load environment variables
load_dotenv()

# Set OpenAI API key
api_key = os.getenv("OPENAI_API_KEY")
if not api_key:
    raise ValueError("API key not found. Please set the OPENAI_API_KEY environment variable.")
os.environ["OPENAI_API_KEY"] = api_key

def load_retrieval_qa_chain():
    # Load embeddings
    embeddings = OpenAIEmbeddings()
    
    # Load vector store
    vectorstore = Chroma(persist_directory="./chroma_db", embedding_function=embeddings)

    # Initialize ChatOpenAI model
    llm = ChatOpenAI(model_name="gpt-4o-mini", temperature=0)  # "gpt-4o-mini

    # Create ConversationalRetrievalChain
    qa_chain = ConversationalRetrievalChain.from_llm(
        llm,
        vectorstore.as_retriever(),
        return_source_documents=True
    )

    return qa_chain

def extract_text_from_pdf(file_path):
    documents = []
    with pdfplumber.open(file_path) as pdf:
        for page_num, page in enumerate(pdf.pages):
            text = page.extract_text()
            if text:
                # Split text into chunks
                text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
                chunks = text_splitter.split_text(text)
                for chunk in chunks:
                    doc = Document(page_content=chunk, metadata={"source": os.path.basename(file_path), "page": page_num + 1})
                    documents.append(doc)
    return documents

def embed_documents():
    embeddings = OpenAIEmbeddings(api_key=os.getenv("OPENAI_API_KEY"))
    vectorstore = Chroma(persist_directory="./chroma_db", embedding_function=embeddings)
    
    pdf_files = [f for f in os.listdir("./documents") if f.endswith('.pdf')]
    documents = []
    with ThreadPoolExecutor() as executor:
        results = executor.map(extract_text_from_pdf, [f"./documents/{pdf_file}" for pdf_file in pdf_files])
        for result in results:
            documents.extend(result)
    vectorstore.add_documents(documents)

def update_embeddings():
    embeddings = OpenAIEmbeddings(api_key=os.getenv("OPENAI_API_KEY"))
    vectorstore = Chroma(persist_directory="./chroma_db", embedding_function=embeddings)
    
    # Retrieve existing documents
    existing_files = set()
    for doc in vectorstore.similarity_search(""):
        existing_files.add(doc.metadata["source"])
    
    pdf_files = [f for f in os.listdir("./documents") if f.endswith('.pdf')]
    
    new_files = [f for f in pdf_files if f not in existing_files]
    documents = []
    with ThreadPoolExecutor() as executor:
        results = executor.map(extract_text_from_pdf, [f"./documents/{pdf_file}" for pdf_file in new_files])
        for result in results:
            documents.extend(result)
    vectorstore.add_documents(documents)

# Generate answer for a query
def get_answer(qa_chain, query, chat_history):
    formatted_history = [(q, a) for q, a in zip(chat_history[::2], chat_history[1::2])]
    
    response = qa_chain.invoke({"question": query, "chat_history": formatted_history})
    
    answer = response["answer"]
    
    source_docs = response.get("source_documents", [])
    source_texts = [f"{os.path.basename(doc.metadata['source'])} (Page {doc.metadata['page']})" for doc in source_docs]
    
    return {"answer": answer, "sources": source_texts}

# Example usage
if __name__ == "__main__":
    update_embeddings()  # Update embeddings with new documents
    qa_chain = load_retrieval_qa_chain()
    question = """당신은 RAG(Retrieval-Augmented Generation) 기반 AI 어시스턴트입니다. 다음 지침을 따라 사용자 질문에 답하세요:

1. 검색 결과 활용: 제공된 검색 결과를 분석하고 관련 정보를 사용해 답변하세요.

2. 정확성 유지: 정보의 정확성을 확인하고, 불확실한 경우 이를 명시하세요.

3. 간결한 응답: 질문에 직접 답하고 핵심 내용에 집중하세요.

4. 추가 정보 제안: 관련된 추가 정보가 있다면 언급하세요.

5. 윤리성 고려: 객관적이고 중립적인 태도를 유지하세요.

6. 한계 인정: 답변할 수 없는 경우 솔직히 인정하세요.

7. 대화 유지: 자연스럽게 대화를 이어가고, 필요시 후속 질문을 제안하세요.
항상 정확하고 유용한 정보를 제공하는 것을 목표로 하세요."""

    response = get_answer(qa_chain, question, [])
    print(f"Question: {question}")
    print(f"Answer: {response['answer']}")
    print(f"Sources: {response['sources']}")