main.py

from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Chroma
from langchain_google_genai import GoogleGenerativeAIEmbeddings, ChatGoogleGenerativeAI
import PyPDF2
import os
import gradio as gr
import google.generativeai as genai
from langchain.chains import ConversationalRetrievalChain
from langchain_huggingface import HuggingFaceEmbeddings
from deep_translator import GoogleTranslator

print('程式初始化')

# 設定 Google API 金鑰
genai.configure(api_key=os.getenv("GOOGLE_API_KEY"))

# 選擇模型
llm_model = 'gemini-1.5-flash'

embeddings_model = "models/embedding-001"
pdf_dir = 'data'

# 讀取 PDF 檔案
print('-' * 21, '讀取資料', '-' * 21)
docs = ""
for filename in os.listdir(pdf_dir):
    if filename.endswith('.pdf'):
        try:
            with open(os.path.join(pdf_dir, filename), 'rb') as pdf_file:
                pdf_reader = PyPDF2.PdfReader(pdf_file)
                for i in range(len(pdf_reader.pages)):  
                    page = pdf_reader.pages[i]  
                    docs += page.extract_text()
                print('讀取成功：',filename)
        except:
            print('讀取失敗：',filename)
            

print('-' * 21, '讀取完成', '-' * 21)
# 分割文本
if docs:
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=300, chunk_overlap=50)
    texts = text_splitter.split_text(docs)

    # 建立嵌入模型和檢索器
    embeddings = GoogleGenerativeAIEmbeddings(
        model=embeddings_model, google_api_key=os.getenv("GOOGLE_API_KEY")
    )
    retriever = Chroma.from_texts(texts, embeddings).as_retriever(search_kwargs={"k": 5})
    print('分割文本完成')

    # 初始化 Gemini 模型
    llm = ChatGoogleGenerativeAI(
        model=llm_model, temperature=0, google_api_key=os.getenv("GOOGLE_API_KEY")
    )
    print('模型載入完成')


# 定義翻譯函數
def translate_to_english(text):
    return GoogleTranslator(source='auto', target='en').translate(text)

def translate_to_chinese(text):
    return GoogleTranslator(source='auto', target='zh-TW').translate(text)


# 定義 invoke 函數
# 初始化 chat_history 为空
chat_history = []

def invoke(question):
    print('invoke 函數觸發')
    if docs:
        system_prompt = (
            "You are an assistant for question-answering tasks. "
            "Use the following pieces of retrieved context to answer the question. "
            
        )
        #"If you don't know the answer, say that you don't know."
        # 初始化 ConversationalRetrievalChain
        qa_chain = ConversationalRetrievalChain.from_llm(
            llm=llm, retriever=retriever
        )

        # 调用链并传递 chat_history
        question = translate_to_english(question)
        # chat_history = translate_to_english(chat_history)
        response = qa_chain.invoke({"question": question, "chat_history": chat_history})
        # response = qa_chain.invoke({"question": question})
        response = translate_to_chinese(response['answer'])

        # 更新 chat_history，保留上下文
        # chat_history += question
        # chat_history += response
    else:
        response = 'No context!'
    
    return response


# Gradio 介面配置
description = "Gradio UI using the Gemini-1.5-Flash model for RAG."
print('description')
gr.close_all()

demo = gr.Interface(
    fn=invoke,
    inputs=gr.Textbox(label="Question", lines=5),
    outputs=gr.Textbox(label="Response", lines=5),
    title="Gemini-RAG",
    description=description
)

demo.launch(share=True)