Update app.py

app.py

@@ -9,10 +9,8 @@ from transformers.models.llama.configuration_llama import LlamaConfig
 from huggingface_hub import hf_hub_download
 import gradio as gr
 
-# --- 解決 Matplotlib 權限問題 ---
-os.environ["MPLCONFIGDIR"] = "/tmp/matplotlib"
-
 # === Enterprise Environment Settings ===
+# Redirect cache directories to writable locations (e.g., /tmp)
 os.environ["HOME"] = "/tmp"
 os.environ["XDG_CACHE_HOME"] = "/tmp/.cache"
 os.environ["HF_HOME"] = "/tmp/huggingface"
@@ -20,9 +18,10 @@ os.environ["TRANSFORMERS_CACHE"] = "/tmp/huggingface/transformers"
 os.environ["HF_DATASETS_CACHE"] = "/tmp/huggingface/datasets"
 os.environ["HF_METRICS_CACHE"] = "/tmp/huggingface/metrics"
 os.environ["GRADIO_FLAGGING_DIR"] = "/tmp/flagged"
+
+# Set SentenceTransformers and HF_HUB cache directories to avoid writing to system root
 os.environ["SENTENCE_TRANSFORMERS_HOME"] = "/tmp/sentence_transformers"
 os.environ["HF_HUB_CACHE"] = "/tmp/huggingface/hf_cache"
-os.environ["HF_HUB_DOWNLOAD_TIMEOUT"] = "60"
 
 # === Load Required Modules ===
 from langchain.embeddings import HuggingFaceEmbeddings
@@ -35,15 +34,20 @@ from langchain.chains import ConversationalRetrievalChain
 from langchain.memory import ConversationBufferMemory
 from langchain_community.document_loaders import PyPDFLoader, TextLoader, UnstructuredWordDocumentLoader
 from langchain.text_splitter import RecursiveCharacterTextSplitter
-from langchain.chains.summarize import load_summarize_chain
 from tempfile import mkdtemp
+
+# Import Summarization Chain (Map-Reduce)
+from langchain.chains.summarize import load_summarize_chain
+
+# Import pandas for CSV handling
 import pandas as pd
 
 # === Multi-Agent Imports ===
+from langchain.agents import initialize_agent, Tool
+from langchain.agents.agent_types import AgentType
+
+# Import SerpAPI (real external search integration)
 from serpapi import GoogleSearch
-# CrewAI 部分：完全使用 CrewAI 的 Agent、Task、Crew 與 @tool 裝飾器
-from crewai import Crew, Agent, Task, Process
-from crewai.tools import tool
 
 # === Model and Device Setup ===
 if torch.backends.mps.is_available():
@@ -54,10 +58,14 @@ else:
     device = "cpu"
 print(f"Using device => {device}")
 
-hf_token = os.environ.get("HF_TOKEN")
-openai_api_key = os.environ.get("OPENAI_API_KEY")
+# Ensure environment variables are set for API keys:
+hf_token = os.environ.get("HF_TOKEN")  # Hugging Face access token
+openai_api_key = os.environ.get("OPENAI_API_KEY")  # OpenAI API key
+# SERPAPI_API_KEY must be set for external search
+
 model_id = "ChienChung/my-llama-1b"
 
+# Download and load model config
 config_path = hf_hub_download(
     repo_id=model_id,
     filename="config.json",
@@ -66,8 +74,13 @@ config_path = hf_hub_download(
 )
 with open(config_path, "r", encoding="utf-8") as f:
     config_dict = json.load(f)
+
 if "rope_scaling" in config_dict:
-    config_dict["rope_scaling"] = {"type": "dynamic", "factor": config_dict["rope_scaling"].get("factor", 32.0)}
+    config_dict["rope_scaling"] = {
+        "type": "dynamic",
+        "factor": config_dict["rope_scaling"].get("factor", 32.0)
+    }
+
 model_config = LlamaConfig.from_dict(config_dict)
 model_config.trust_remote_code = True
 
@@ -93,6 +106,7 @@ if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token
 print("Tokenizer loaded!")
 
+# Build query pipeline
 query_pipeline = transformers.pipeline(
     "text-generation",
     model=model,
@@ -109,6 +123,7 @@ query_pipeline = transformers.pipeline(
 print("Loading Chroma DB for Biden Speech...")
 if not os.path.exists("/tmp/chroma_db"):
     shutil.copytree("./chroma_db", "/tmp/chroma_db")
+
 embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")
 vectordb = Chroma(persist_directory="/tmp/chroma_db", embedding_function=embeddings)
 retriever = vectordb.as_retriever()
@@ -143,22 +158,9 @@ qa_gpt = ConversationalRetrievalChain.from_llm(
     combine_docs_chain_kwargs={"prompt": custom_prompt}
 )
 
-# === Helper Function：從上傳文件中提取檔案路徑 ===
-def get_file_path(file):
-    if isinstance(file, str):
-        return file
-    elif isinstance(file, dict):
-        # 優先使用 "data" 鍵，其次是 "name"
-        return file.get("data", file.get("name", None))
-    elif hasattr(file, "save"):
-        temp_dir = mkdtemp()
-        file_path = os.path.join(temp_dir, file.name)
-        file.save(file_path)
-        return file_path
-    else:
-        return None
+# === Function Definitions ===
 
-# === 原有功能（Tab 1～Tab 4）函式 ===
+# Biden Q&A (LLaMA)
 def rag_llama_qa(query):
     output = RetrievalQA.from_chain_type(
         llm=llm_local,
@@ -171,19 +173,32 @@ def rag_llama_qa(query):
     idx = lower_text.find("answer:")
     return output[idx + len("answer:"):].strip() if idx != -1 else output
 
+# Biden Q&A (GPT-4)
 def rag_gpt4_qa(query):
     return qa_gpt.run(query)
 
+# Document Q&A (Upload + GPT-4)
 def upload_and_chat(file, query):
-    file_path = get_file_path(file)
-    if file_path is None:
-        return "Unable to obtain the uploaded file path."
+    if isinstance(file, str):
+        file_path = file
+    elif isinstance(file, dict):
+        file_path = file.get("name", None)
+        if file_path is None:
+            return "Unable to obtain the uploaded file path."
+    elif hasattr(file, "save"):
+        temp_dir = mkdtemp()
+        file_path = os.path.join(temp_dir, file.name)
+        file.save(file_path)
+    else:
+        return "Unable to process the file format."
+
     if file_path.lower().endswith(".pdf"):
         loader = PyPDFLoader(file_path)
     elif file_path.lower().endswith(".docx"):
         loader = UnstructuredWordDocumentLoader(file_path)
     else:
         loader = TextLoader(file_path)
+
     docs = loader.load()
     chunks = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50).split_documents(docs)
     db = FAISS.from_documents(chunks, embeddings)
@@ -197,63 +212,96 @@ def upload_and_chat(file, query):
     )
     return qa_temp.run(query)
 
+def patched_combine_llm_outputs(self, llm_outputs):
+    return {}
+ChatOpenAI._combine_llm_outputs = patched_combine_llm_outputs
+
 initial_prompt = PromptTemplate(
     input_variables=["text"],
-    template="""Write a concise and structured summary of the following content. Focus on capturing the main ideas and key details:
+    template="""
+Write a concise summary of the following text:
 
-{text}
+"{text}"
 
---- Summary ---
+CONCISE SUMMARY:
 """
 )
+
 refine_prompt = PromptTemplate(
     input_variables=["existing_answer", "text"],
-    template="""You already have an existing summary:
-{existing_answer}
-
-Refine the summary based on the new content below. Add or update information only if it's relevant. Keep it concise:
+    template="""
+You have a summary already:
+"{existing_answer}"
 
-{text}
+Now refine the summary with the following text:
+"{text}"
 
---- Refined Summary ---
+REVISED SUMMARY:
 """
 )
-
+# Document Summarization (Refine)
 def document_summarize(file):
-    file_path = get_file_path(file)
-    if file_path is None:
-        return "Unable to obtain the uploaded file."
+    if isinstance(file, str):
+        file_path = file
+    elif isinstance(file, dict):
+        file_path = file.get("name", None)
+        if file_path is None:
+            return "Unable to obtain the uploaded file."
+    elif hasattr(file, "save"):
+        temp_dir = mkdtemp()
+        file_path = os.path.join(temp_dir, file.name)
+        file.save(file_path)
+    else:
+        return "Unable to process the file format."
+
     if file_path.lower().endswith(".pdf"):
         loader = PyPDFLoader(file_path)
     elif file_path.lower().endswith(".docx"):
         loader = UnstructuredWordDocumentLoader(file_path)
     else:
         loader = TextLoader(file_path)
+
     docs = loader.load()
     summarize_chain = load_summarize_chain(llm_gpt4, chain_type="refine", question_prompt=initial_prompt, refine_prompt=refine_prompt)
     summary = summarize_chain.invoke(docs)
     return summary['output_text']
 
+# CSVAgent: Process CSV queries for financial calculations and data analysis
 def csv_agent(file, query):
-    file_path = get_file_path(file)
-    if file_path is None:
-        return "Unable to obtain the uploaded CSV file."
+    if isinstance(file, str):
+        file_path = file
+    elif isinstance(file, dict):
+        file_path = file.get("name", None)
+        if file_path is None:
+            return "Unable to obtain the uploaded CSV file."
+    elif hasattr(file, "save"):
+        temp_dir = mkdtemp()
+        file_path = os.path.join(temp_dir, file.name)
+        file.save(file_path)
+    else:
+        return "Unable to process the file format."
     try:
         df = pd.read_csv(file_path)
     except Exception as e:
-        return f"Error reading CSV: {e}"
+        return f"Error reading CSV: {str(e)}"
     safe_dict = {"df": df, "pd": pd}
     try:
         result = eval(query, {"__builtins__": {}}, safe_dict)
         return str(result)
     except Exception as e:
-        return f"Query error: {e}"
+        return f"Query error: {str(e)}"
 
+# SearchAgent: Use SerpAPI to perform real external search
 def search_agent(query):
     api_key = os.environ.get("SERPAPI_API_KEY")
     if not api_key:
         return "SERPAPI_API_KEY not set. Please set the environment variable."
-    params = {"engine": "google", "q": query, "api_key": api_key, "num": 5}
+    params = {
+        "engine": "google",
+        "q": query,
+        "api_key": api_key,
+        "num": 5  # Adjust number of results as needed
+    }
     search = GoogleSearch(params)
     results = search.get_dict()
     if "organic_results" in results:
@@ -267,10 +315,41 @@ def search_agent(query):
     else:
         return "No results found."
 
+# CrewAI Agent: Simulate CrewAI response integrating chat history and tool routing
+def crew_ai_agent(query):
+    return f"CrewAI response for '{query}' with integrated chat history."
+
+
+# === Global State for Tab 5 ===
+session_retriever = None
+session_qa_chain = None
+
+def save_file(file):
+    if isinstance(file, str):
+        return file
+    elif isinstance(file, dict):
+        return file.get("name", None)
+    elif hasattr(file, "save"):
+        temp_dir = mkdtemp()
+        file_path = os.path.join(temp_dir, file.name)
+        file.save(file_path)
+        return file_path
+    else:
+        return None
+# === 新增 Uploaded QA 函式與 Tool (不更動其他代碼) ===
 def uploaded_qa(file, query):
-    file_path = get_file_path(file)
-    if file_path is None:
-        return "Unable to obtain the uploaded file path."
+    if isinstance(file, str):
+        file_path = file
+    elif isinstance(file, dict):
+        file_path = file.get("name", None)
+        if file_path is None:
+            return "Unable to obtain the uploaded file path."
+    elif hasattr(file, "save"):
+        temp_dir = mkdtemp()
+        file_path = os.path.join(temp_dir, file.name)
+        file.save(file_path)
+    else:
+        return "Unable to process the file format."
     if file_path.lower().endswith(".pdf"):
         loader = PyPDFLoader(file_path)
     elif file_path.lower().endswith(".docx"):
@@ -290,115 +369,46 @@ def uploaded_qa(file, query):
     )
     return qa_temp.run(query)
 
-# === CrewAI Multi-Agent System (Tab 5) ===
-# 完全捨棄 langchain.agents.Tool，使用 CrewAI 的 @tool 裝飾器來定義工具
-from pydantic import BaseModel
-class SimpleQuery(BaseModel):
-    query: str
-
-@tool("summarise")
-def summarise_tool(query: str) -> str:
-    """Summarise: 使用文件摘要功能。"""
-    global session_retriever, session_qa_chain
-    if session_retriever is None:
-        return "尚未上傳文件。"
-    try:
-        docs = session_retriever.get_relevant_documents(query if query.strip() else "summary")
-        if not docs:
-            return "無法從文件中找到相關內容。"
-        summarize_chain = load_summarize_chain(llm_gpt4, chain_type="refine", question_prompt=initial_prompt, refine_prompt=refine_prompt)
-        summary = summarize_chain.invoke(docs)
-        return summary['output_text']
-    except Exception as e:
-        return f"摘要錯誤: {e}"
-
-@tool("python_calc")
-def python_calc_tool(query: str) -> str:
-    """Python Calculation: 執行簡單計算。"""
-    try:
-        return str(eval(query))
-    except Exception as e:
-        return f"計算錯誤: {e}"
-
-@tool("search_agent")
-def search_tool_func(query: str) -> str:
-    """Search: 執行網路搜尋。"""
-    return search_agent(query)
-
-@tool("uploaded_qa")
-def uploaded_qa_tool_func(query: str) -> str:
-    """Document QA: 根據上傳文件回答問題。"""
-    global session_qa_chain
-    if session_qa_chain is not None:
-        try:
-            return session_qa_chain.run(query)
-        except Exception as e:
-            return f"文檔問答錯誤: {e}"
-    else:
-        return "尚未上傳文件。"
-
-# 建立 CrewAI 代理（僅針對 Tab 5）
-summarizer_agent = Agent(
-    role="文件摘要助手",
-    goal="對上傳文件內容進行摘要",
-    backstory="你是一位專業的摘要專家，能抓住長文的重點。",
-    tools=[summarise_tool],
-    verbose=True
-)
-document_qa_agent = Agent(
-    role="文件問答專家",
-    goal="根據上傳文件回答問題",
-    backstory="你精通文檔內容，能從中找出問題答案。",
-    tools=[uploaded_qa_tool_func],
-    verbose=True
-)
-general_agent = Agent(
-    role="綜合助手",
-    goal="回答一般問題，執行計算與網路搜尋",
-    backstory="你是一位多才多藝的AI助理，能根據需要使用工具。",
-    tools=[python_calc_tool, search_tool_func],
-    verbose=True
-)
+uploaded_qa_tool = Tool(name="uploaded_qa", func=uploaded_qa, description="Tool for answering questions based on uploaded documents without conflicting with Biden QA.")
 
-router_task = Task(
-    description="根據使用者查詢自動決定使用哪個工具進行回答。",
-    expected_output="最終回答",
-    agent=general_agent
-)
+# === Multi-Agent Tools and Initialization ===
+faq_tool = Tool(name="faq_qa", func=qa_gpt.run, description="Document Q&A tool that answers questions based on document content.")
+summarise_tool = Tool(name="summarise", func=lambda q: llm_gpt4.predict(f"Summarise this:\n{q}"), description="Tool for summarising long content.")
+data_tool = Tool(name="python_calc", func=lambda q: str(eval(q)), description="Execute Python-based calculations.")
+csv_tool = Tool(name="csv_agent", func=csv_agent, description="CSV file query and data analysis tool.")
+search_tool = Tool(name="search_agent", func=search_agent, description="External search tool using SerpAPI.")
+crew_tool = Tool(name="crew_ai", func=crew_ai_agent, description="CrewAI agent integrating chat history and tool routing.")
 
-crew = Crew(
-    agents=[summarizer_agent, document_qa_agent, general_agent],
-    tasks=[router_task],
-    process=Process.sequential,
-    verbose=True
-)
+tools = [summarise_tool, data_tool, csv_tool, search_tool, crew_tool, uploaded_qa_tool]
+agent_executor = initialize_agent(tools, llm=llm_gpt4, agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose=True)
 
-def multi_agent_chat(query: str) -> str:
+def multi_agent_chat(query):
     try:
-        return crew.run(query)
+        return agent_executor.run(query)
     except Exception as e:
-        return f"Error: {e}"
+        return f"Error: {str(e)}"
 
-def multi_agent_chat_advanced(query: str, file=None) -> str:
+# === 統一上傳檔案並由後端自動判斷功能的 Advanced Multi-Agent 功能 ===
+def multi_agent_chat_advanced(query, file=None):
     global session_retriever, session_qa_chain
-    # 定義一些明顯與文件無關的關鍵字
-    non_doc_keywords = ["calculate", "sum", "date", "time", "how many", "how much", "weather", "temperature"]
-    use_file_chain = True
-    for kw in non_doc_keywords:
-        if kw in query.lower():
-            use_file_chain = False
-            break
 
     if file is not None:
-        file_path = get_file_path(file)
+        # Step 1: 處理上傳並建立 retriever + QA chain
+        file_path = save_file(file)
         if file_path is None:
-            return "Unable to process the file format."
+            return "無法處理該檔案格式。"
+
         if file_path.lower().endswith(".csv"):
             return csv_agent(file, query)
+
         elif file_path.lower().endswith((".pdf", ".txt", ".docx")):
-            loader = (PyPDFLoader(file_path) if file_path.lower().endswith(".pdf")
-                      else UnstructuredWordDocumentLoader(file_path) if file_path.lower().endswith(".docx")
-                      else TextLoader(file_path))
+            loader = (
+                PyPDFLoader(file_path)
+                if file_path.endswith(".pdf")
+                else UnstructuredWordDocumentLoader(file_path)
+                if file_path.endswith(".docx")
+                else TextLoader(file_path)
+            )
             docs = loader.load()
             chunks = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50).split_documents(docs)
             db = FAISS.from_documents(chunks, embeddings)
@@ -409,30 +419,23 @@ def multi_agent_chat_advanced(query: str, file=None) -> str:
                 memory=ConversationBufferMemory(memory_key="chat_history", return_messages=True),
                 combine_docs_chain_kwargs={"prompt": custom_prompt}
             )
-            if use_file_chain:
-                return session_qa_chain.run(query)
+
+            # 如果 query 是摘要
+            if any(kw in query.lower() for kw in ["summarize", "summary", "摘要", "總結"]):
+                return document_summarize(file)
             else:
-                try:
-                    return crew.run(query)
-                except Exception as e:
-                    return f"Error: {e}"
+                return session_qa_chain.run(query)
         else:
-            return "Unsupported file format."
+            return "不支援的檔案格式。"
+
+    # Step 2: 已經上傳過檔案 → 使用現成的 retriever
     elif session_qa_chain is not None:
-        if use_file_chain:
-            return session_qa_chain.run(query)
-        else:
-            try:
-                return crew.run(query)
-            except Exception as e:
-                return f"Error: {e}"
+        return session_qa_chain.run(query)
     else:
-        try:
-            return crew.run(query)
-        except Exception as e:
-            return f"Error: {e}"
+        return "請先上傳檔案，再進行提問。"
+
+# === Gradio Interface Settings (All texts in English) ===
 
-# === Gradio Interface Settings ===
 demo_description = """
 **Context**:
 This demo uses a Retrieval-Augmented Generation (RAG) system based on 
@@ -451,6 +454,7 @@ If no relevant information is found in the document, the system will say "No rel
 
 Feel free to ask any question related to Biden’s 2023 State of the Union Address.
 """
+
 demo_description2 = """
 **Context**:
 This demo uses a Retrieval-Augmented Generation (RAG) system based on 
@@ -469,13 +473,19 @@ If no relevant information is found in the document, the system will say "No rel
 
 Feel free to ask any question related to Biden’s 2023 State of the Union Address.
 """
+
 demo_description3 = """
 **Context**:
-Upload a PDF, TXT, or DOCX file and ask a question about its content.
-This demo uses GPT-4 to answer questions based on the content of your uploaded document.
+This demo uses a Retrieval-Augmented Generation (RAG) system based on 
+your uploaded document. 
+All responses are grounded in this document. 
+If no relevant information is found in the document, the system will say "No relevant info found."
+
+*Note: The GPT module supports follow-up questions with conversation history management, enabling more interactive and context-aware discussions.*
 
 Feel free to ask any question related to your document.
 """
+
 demo_description4 = """
 **Context**:
 This assistant performs multi-agent tasks using tools such as:
@@ -484,9 +494,10 @@ This assistant performs multi-agent tasks using tools such as:
 - Financial or CSV-style logic queries
 - Multi-step reasoning via agent orchestration
 
-The system will automatically select the appropriate function based on the uploaded file (CSV, PDF, TXT, DOCX) and the query content.
-For example, if the query contains "summarize"/"摘要", it will summarize the document; if it's CSV data, it will perform data analysis.
+系統會根據上傳檔案（CSV、PDF、TXT、DOCX）以及查詢內容，自動選擇呼叫適當的功能，
+例如：若查詢中含有 "summarize"/"摘要"，則對文件進行摘要；若是 CSV 則進行資料查詢等。
 """
+
 demo_description5 = """
 **Context**:
 This demo uses Document Summarization via a Map-Reduce chain.