import gradio as gr import os from langchain.chains import ConversationalRetrievalChain from langchain.text_splitter import CharacterTextSplitter from langchain_community.document_loaders import PyPDFLoader, Docx2txtLoader, TextLoader from langchain_community.vectorstores import Chroma from langchain_openai import ChatOpenAI, OpenAIEmbeddings from dotenv import load_dotenv # 加載環境變量 load_dotenv() os.environ["OPENAI_API_KEY"] = os.getenv("OPENAI_API_KEY") # 驗證 OpenAI API Key api_key = os.getenv('OPENAI_API_KEY') if not api_key: raise ValueError("請設置 'OPENAI_API_KEY' 環境變數") # OpenAI API key openai_api_key = api_key # 將聊天歷史轉換為適合 LangChain 的二元組格式 def transform_history_for_langchain(history): return [(chat[0], chat[1]) for chat in history if chat[0]] # 使用整數索引來訪問元組中的元素 # 將 Gradio 的歷史紀錄轉換為 OpenAI 格式 def transform_history_for_openai(history): new_history = [] for chat in history: if chat[0]: new_history.append({"role": "user", "content": chat[0]}) if chat[1]: new_history.append({"role": "assistant", "content": chat[1]}) return new_history # 載入和處理文件的函數 def load_and_process_documents(folder_path): documents = [] for file in os.listdir(folder_path): file_path = os.path.join(folder_path, file) if file.endswith(".pdf"): loader = PyPDFLoader(file_path) documents.extend(loader.load()) elif file.endswith('.docx') or file.endswith('.doc'): loader = Docx2txtLoader(file_path) documents.extend(loader.load()) elif file.endswith('.txt'): loader = TextLoader(file_path) documents.extend(loader.load()) text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=10) documents = text_splitter.split_documents(documents) vectordb = Chroma.from_documents( documents, embedding=OpenAIEmbeddings(), persist_directory="./tmp" ) return vectordb # 初始化向量數據庫為全局變量 if 'vectordb' not in globals(): vectordb = load_and_process_documents("./") # 定義查詢處理函數 def handle_query(user_message, temperature, chat_history): try: if not user_message: return chat_history # 返回不變的聊天記錄 # 使用 LangChain 的 ConversationalRetrievalChain 處理查詢 preface = """ 指令: 全部以繁體中文呈現,200字以內。 除了與文件相關內容可回答之外,與文件內容不相關的問題都必須回答:這問題很深奧,需要請示JohnLiao大神... """ query = f"{preface} 查詢內容:{user_message}" # 提取之前的回答作為上下文,並轉換成 LangChain 支持的格式 previous_answers = transform_history_for_langchain(chat_history) pdf_qa = ConversationalRetrievalChain.from_llm( ChatOpenAI(temperature=temperature, model_name='gpt-4'), retriever=vectordb.as_retriever(search_kwargs={'k': 6}), return_source_documents=True, verbose=False ) # 調用模型進行查詢 result = pdf_qa.invoke({"question": query, "chat_history": previous_answers}) # 確保 'answer' 在結果中 if "answer" not in result: return chat_history + [("系統", "抱歉,出現了一個錯誤。")] # 更新對話歷史中的 AI 回應 chat_history[-1] = (user_message, result["answer"]) # 更新最後一個記錄,配對用戶輸入和 AI 回應 return chat_history except Exception as e: return chat_history + [("系統", f"出現錯誤: {str(e)}")] # 使用 Gradio 的 Blocks API 創建自訂聊天介面 with gr.Blocks() as demo: gr.Markdown("

AI 小助教

") chatbot = gr.Chatbot() state = gr.State([]) with gr.Row(): with gr.Column(scale=0.85): txt = gr.Textbox(show_label=False, placeholder="請輸入您的問題...") with gr.Column(scale=0.15, min_width=0): submit_btn = gr.Button("提問") # 用戶輸入後立即顯示提問文字,不添加回應部分,並清空輸入框 def user_input(user_message, history): history.append((user_message, "")) # 顯示提問文字,回應部分為空字符串 return history, "", history # 返回清空的輸入框以及更新的聊天歷史 # 處理 AI 回應,更新回應部分 def bot_response(history): user_message = history[-1][0] # 獲取最新的用戶輸入 history = handle_query(user_message, 0.7, history) # 調用處理函數 return history, history # 返回更新後的聊天記錄 # 先顯示提問文字,然後處理 AI 回應,並清空輸入框 submit_btn.click(user_input, [txt, state], [chatbot, txt, state], queue=False).then( bot_response, state, [chatbot, state] ) # 支援按 "Enter" 提交問題,立即顯示提問文字並清空輸入框 txt.submit(user_input, [txt, state], [chatbot, txt, state], queue=False).then( bot_response, state, [chatbot, state] ) # 啟動 Gradio 應用 demo.launch()