Update app.py

app.py

@@ -6,11 +6,12 @@ import torch
 import re
 import transformers
 import chardet
-import pandas as pd
-from datetime import datetime
-from tempfile import mkdtemp
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers.models.llama.configuration_llama import LlamaConfig
+from huggingface_hub import hf_hub_download
+import gradio as gr
 
-# --- Fix Matplotlib permissions ---
+# --- 解決 Matplotlib 權限問題 ---
 os.environ["MPLCONFIGDIR"] = "/tmp/matplotlib"
 
 # === Enterprise Environment Settings ===
@@ -25,36 +26,34 @@ os.environ["SENTENCE_TRANSFORMERS_HOME"] = "/tmp/sentence_transformers"
 os.environ["HF_HUB_CACHE"] = "/tmp/huggingface/hf_cache"
 os.environ["HF_HUB_DOWNLOAD_TIMEOUT"] = "60"
 
-# === Updated Imports from langchain-community ===
-from langchain_community.embeddings import HuggingFaceEmbeddings
-from langchain_community.vectorstores import Chroma, FAISS
+# === Load Required Modules ===
+from langchain.embeddings import HuggingFaceEmbeddings
+from langchain.vectorstores import Chroma, FAISS
 from langchain.chains import RetrievalQA
 from langchain.prompts import PromptTemplate
-from langchain_community.llms import HuggingFacePipeline
-from langchain_community.chat_models import ChatOpenAI
+from langchain.llms import HuggingFacePipeline
+from langchain.chat_models import ChatOpenAI
 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
 from langchain.schema import AIMessage
+from datetime import datetime
+import numexpr as ne
+import pandas as pd
 
-# === Other Imports ===
+# === Multi-Agent Imports ===
 from serpapi import GoogleSearch
-
-# CrewAI Imports
+# 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
 from langchain_experimental.agents import create_pandas_dataframe_agent
-from pydantic import BaseModel
-
-import numexpr as ne
 
-# Global variables for non‑math functions (unchanged)
 session_retriever = None
 session_qa_chain = None
 csv_dataframe = None  # CSV tool will use this
-
 # === Safe Result Formatter ===
 def safe_format_result(result) -> str:
     try:
@@ -70,7 +69,6 @@ def safe_format_result(result) -> str:
             return str(result)
     except Exception as e:
         return f"Error formatting result: {e}"
-
 # === Model and Device Setup ===
 if torch.backends.mps.is_available():
     device = "mps"
@@ -174,11 +172,12 @@ qa_gpt = ConversationalRetrievalChain.from_llm(
     combine_docs_chain_kwargs={"prompt": custom_prompt}
 )
 
-# === Helper Function: Get file path from uploaded file ===
+# === 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()
@@ -188,7 +187,7 @@ def get_file_path(file):
     else:
         return None
 
-# === Original functionalities (Tabs 1-4) unchanged ===
+# === 原有功能（Tab 1～Tab 4）函式 ===
 def rag_llama_qa(query):
     output = RetrievalQA.from_chain_type(
         llm=llm_local,
@@ -264,6 +263,7 @@ def document_summarize(file):
     summary = summarize_chain.invoke(docs)
     return summary['output_text']
 
+
 def csv_agent(file, query):
     file_path = get_file_path(file)
     if file_path is None:
@@ -323,130 +323,418 @@ def uploaded_qa(file, query):
     )
     return qa_temp.run(query)
 
-# === Math Natural Language Tool ===
-def preprocess_math_query(query: str) -> str:
-    """
-    Convert expressions like '10% of 800' to '0.10 * 800'
-    """
-    pattern = r"(\d+(?:\.\d+)?)%\s+of\s+(\d+(?:\.\d+)?)"
-    match = re.search(pattern, query)
-    if match:
-        perc = float(match.group(1)) / 100.0
-        number = match.group(2)
-        return f"{perc} * {number}"
-    return query
-
-def convert_trig(expr: str) -> str:
-    """
-    Convert sine functions from degrees to radians.
-    For example: sin(45) -> sin(pi*45/180)
-    """
-    pattern = r"sin\(\s*(\d+(?:\.\d+)?)\s*\)"
-    return re.sub(pattern, lambda m: f"sin(pi*{m.group(1)}/180)", expr)
-
-def convert_nl_to_expr(query: str) -> str:
-    """
-    Use GPT-4 to convert a natural language math question into a valid mathematical expression.
-    """
-    prompt = (
-        "Convert the following natural language math question into a valid mathematical expression "
-        "using standard arithmetic symbols. Output only the expression.\n\n"
-        "Question: " + query
-    )
-    response = llm_gpt4.invoke(prompt)
-    if isinstance(response, AIMessage):
-        response = response.content
-    return response.strip()
-
-def generate_nl_explanation(expression: str, result: str) -> str:
-    """
-    Use GPT-4 to generate a natural language explanation for the evaluated expression.
-    """
-    prompt = (
-        f"Explain in a clear, human-friendly sentence the result of the expression '{expression}' which evaluates to '{result}'."
-    )
-    explanation = llm_gpt4.invoke(prompt)
-    if isinstance(explanation, AIMessage):
-        explanation = explanation.content
-    return explanation.strip()
+# === 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
 
-def math_nl_tool_impl(query: str) -> str:
+def _general_chat(query: str) -> str:
     try:
-        # Step 1: Convert natural language to a math expression
-        expression = convert_nl_to_expr(query)
-        # Step 2: Preprocess percentage expressions if any
-        if "%" in expression and "of" in expression:
-            expression = preprocess_math_query(expression)
-        # Step 3: Convert trigonometric functions (assume sine in degrees)
-        expression = convert_trig(expression)
-        # Step 4: Evaluate the expression using numexpr
-        evaluated_result = ne.evaluate(expression)
-        # Step 5: Generate natural language explanation
-        nl_explanation = generate_nl_explanation(expression, str(evaluated_result))
-        return nl_explanation
+        response = llm_gpt4.invoke(query)
+        if isinstance(response, AIMessage):
+            response = response.content  # ✅ 取出真實字串
+        if any(kw in response.lower() for kw in ["i'm not sure", "i don't know", "no information", "can't find"]):
+            return _search_web_tool(query)
+        return response
     except Exception as e:
-        return f"Math NL Tool error: {e}"
-
-@tool("math_nl_tool")
-def math_nl_tool(query: str) -> str:
-    """
-    Math NL Tool:
-    Answer natural language math questions by converting them into a mathematical expression using GPT-4,
-    evaluating the expression with numexpr, and generating a human-friendly explanation.
-    """
-    return math_nl_tool_impl(query)
-
-# === CrewAI Multi-Agent System (For Math Only) ===
-class SimpleQuery(BaseModel):
-    query: str
+        return f"General chat error: {e}"
+@tool("general_chat")
+def general_chat_tool(query: str) -> str:
+    """General assistant: Answer general questions without relying on documents."""
+    try:
+        response = llm_gpt4.invoke(query)
+        if isinstance(response, AIMessage):
+            response = response.content  # ✅ 取出真實字串
+        if any(kw in response.lower() for kw in ["i'm not sure", "i don't know", "no information", "can't find"]):
+            return search_web(query)
+        return response
+    except Exception as e:
+        return f"General chat error: {e}"
+
+def get_time_tool(query: str) -> str:
+    now = datetime.now()
+    return now.strftime("The current local time is %I:%M %p on %B %d, %Y.")
+    
+@tool("summarise")
+def summarise_tool(query: str) -> str:
+    """Summarise: Use document summarisation functionality."""
+    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}"
+        
+def _calc_tool(query: str) -> str:
+    # 判斷是否是明確的數學算式（只包含數學相關符號）
+    if re.match(r"^[a-zA-Z\d\s\+\-\*\/\%\(\)\.]+$", query.strip()):
+        try:
+            result = ne.evaluate(query)
+            return f"The result of the calculation is: {result}"
+        except Exception as e:
+            return f"Error in formula calculation: {e}"
+    # 否則交給 GPT-4o 處理自然語言的問題 + 生成自然語言回答
+    try:
+        response = llm_gpt4.invoke(query)
+        if isinstance(response, AIMessage):
+            response = response.content
+        return response
+    except Exception as e:
+        return f"Natural language math error: {e}"
+        
+@tool("python_calc")
+def python_calc_tool(query: str) -> str:
+    """Python Calculation: Perform basic arithmetic or logical operations."""
+    try:
+        result = ne.evaluate(query)
+        return str(result)
+    except Exception as e:
+        return f"計算錯誤: {e}"
+def _search_web_tool(query: str) -> str:
+    return search_web(query)
+@tool("search_tool")
+def search_tool_func(query: str) -> str:
+    """Search: Perform web searches using external search engines."""
+    return search_web(query)
+
+@tool("uploaded_qa")
+def uploaded_qa_tool_func(query: str) -> str:
+    """Document QA: Answer questions based on the uploaded document content."""
+    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。"
+        
+@tool("csv_agent")
+def csv_tool_func(query: str) -> str:
+    """CSV Agent: Use natural language to analyse uploaded CSV files."""
+    global csv_dataframe
+    if csv_dataframe is None:
+        return "No CSV file uploaded."
+    try:
+        agent = create_pandas_dataframe_agent(llm=llm_gpt4, df=csv_dataframe, verbose=True)
+        return agent.run(f"Here is the table:\n{csv_dataframe.head().to_string(index=False)}\n\n{query}")
+    except Exception as e:
+        return f"CSV Agent error: {e}"
+
+# 建立 CrewAI 代理（僅針對 Tab 5）
+general_agent = Agent(
+    role="General Assistant",
+    goal="Respond to any general query that is not related to documents or CSV files.",
+    backstory="You're an intelligent assistant who answers questions about anything general, such as math, dates, or general knowledge.",
+    tools=[general_chat_tool],
+    verbose=True
+)
+summarizer_agent = Agent(
+    role="Document Summarizer",
+    goal="Summarise the content of the uploaded document.",
+    backstory="You are a professional summarisation expert who can identify key points in long documents.",
+    tools=[summarise_tool],
+    verbose=True
+)
+document_qa_agent = Agent(
+    role="Document QA Specialist",
+    goal="Answer questions based on the uploaded document.",
+    backstory="You are an expert in document understanding and can accurately extract answers.",
+    tools=[uploaded_qa_tool_func],
+    verbose=True
+)
+
+search_agent = Agent(
+    role="Search Expert",
+    goal="Search the web and provide relevant information.",
+    backstory="You are an expert at finding relevant information from the internet.",
+    tools=[search_tool_func],
+    verbose=True
+)
 
 math_agent = Agent(
     role="Math Assistant",
-    goal="Perform natural language math calculations.",
-    backstory="You convert natural language math questions to expressions and compute the result.",
-    tools=[math_nl_tool],
+    goal="Perform accurate arithmetic or logical calculations.",
+    backstory="You are a calculator expert skilled at quick computations.",
+    tools=[python_calc_tool],
+    verbose=True
+)
+csv_agent = Agent(
+    role="CSV Analyst",
+    goal="Analyse tabular data and answer questions about the uploaded CSV file.",
+    backstory="You are skilled in interpreting tabular datasets and can extract numerical or logical insights.",
+    tools=[csv_tool_func],
     verbose=True
 )
-
 router_agent = Agent(
     role="Query Router",
-    goal="Route the query to the best tool.",
-    backstory="You analyze the user's query and choose the appropriate agent.",
-    tools=[math_nl_tool],
+    goal="Determine the most suitable agent or tool to handle the user query.",
+    backstory="You are an intelligent query dispatcher that analyses the user's intent and chooses the best AI agent to answer.",
+    tools=[python_calc_tool, search_tool_func, csv_tool_func, uploaded_qa_tool_func, summarise_tool, general_chat_tool],
     verbose=True
 )
-
 router_task = Task(
     description="""
-Based on the user's query, if it is a math question, use the Math Agent.
+Based on the user's query, decide which agent or tool is best suited to handle it:
+- If the query is related to the content of an uploaded file (e.g., 'what is this document about?'), send it to the **Document QA Agent**.
+- If the query contains words like 'summarize', 'summary', or 'main points', use the **Summarizer Agent**.
+- If the query involves numbers, calculations, or logic (e.g., '50 * 23 - 5', 'what is 10% of 800'), send it to the **Math Agent**.
+- If the user uploaded a CSV file and asks about table content, data trends, or uses words like 'data', 'table', 'csv', 'column', or 'row', send it to the **CSV Agent**.
+- If the user asks about current events, trending topics, or online information (e.g., 'What is LangChain?', 'latest news'), send it to the **Search Agent**.
+- If the query is about current date, time, or day of week (e.g., 'what is today\\'s date?', 'what time is it?', 'what day is it?'), send it to the **Search Agent** instead of the General Agent.
+- If the question is general and not related to documents, calculations, CSVs, or the internet (e.g., 'Who are you?', 'Tell me a fun fact'), send it to the **General Agent**.
+- If none of these apply, use your best judgment to choose the most relevant agent.
 """,
-    expected_output="The final answer from the selected agent.",
+    expected_output="The final answer from the selected agent or tool.",
     agent=router_agent,
     input_variables=["query"]
 )
 
-crew_llm = ChatOpenAI(model_name="gpt-4o-mini", temperature=0.2, openai_api_key=openai_api_key)
-
 crew = Crew(
-    agents=[math_agent],
+    agents=[general_agent, summarizer_agent, document_qa_agent, search_agent, math_agent, csv_agent],
     tasks=[router_task],
     process=Process.sequential,
     verbose=True,
     llm=crew_llm
 )
 
-# === Main Function (Math Focus) ===
+#def multi_agent_chat(query: str) -> str:
+#    print(f"Routing query: {query}")
+#    try:
+        # 🔧 補這段，用來轉換 upload file 語意（擺在 kickoff 前）
+#        if file is not None and ("upload file" in query.lower() or "the file" in query.lower()):
+#            query = f"The user uploaded the following file:\n\n{query}"
+
+#        result = crew.kickoff(inputs={"query": query})
+#        result_str = safe_format_result(result)
+#        if any(kw in result_str.lower() for kw in ["i don't know", "no relevant", "no information", "no answer"]) or result_str.strip() == "":
+#            try:
+#                result = search_agent.execute(query)
+#                return safe_format_result(result)
+#            except Exception as e:
+#                return f"[Fallback Search Error] {e}"
+#        return result_str
+#    except Exception as e:
+#        return f"Error: {e}"
+
 def multi_agent_chat_advanced(query: str, file=None) -> str:
-    # For this version, we focus exclusively on math NL tool.
-    return math_nl_tool(query)
-
-# === Gradio Interface Settings (Math Only) ===
-demo = gr.Interface(
-    fn=multi_agent_chat_advanced,
-    inputs=gr.Textbox(label="Enter your math query"),
-    outputs="text",
-    title="Math NL Tool Demo",
-    description="This tool converts natural language math questions into a mathematical expression, evaluates it, and returns a natural language explanation."
+    global session_retriever, session_qa_chain, csv_dataframe
+
+    # === Step 0: Smart routing without needing uploaded files ===
+    lower_query = query.lower()
+
+    
+    if re.match(r"^[a-zA-Z\d\s\+\-\*\/\%\(\)\.]+$", lower_query):
+        return _calc_tool(query)
+
+    date_keywords = ["what date", "today", "what time", "what day", "現在幾點", "今天幾號", "禮拜幾"]
+    if any(k in lower_query for k in date_keywords):
+        return get_time_tool(query)
+
+    general_keywords = ["who are you", "tell me", "what is your name", "what can you do", "fun fact"]
+    if any(k in lower_query for k in general_keywords):
+        return _general_chat(query)
+
+    # === Step 1: 檢查 file 是否存在與格式 ===
+    file_path = get_file_path(file) if file is not None else None
+
+    # === Step 2: 判斷 Query 是否需要走文件處理邏輯 ===
+    non_doc_keywords = ["calculate", "sum", "date", "time", "how many", "how much", "weather", "temperature"]
+    use_file_chain = not any(kw in query.lower() for kw in non_doc_keywords)
+
+    # === Step 3: 有上傳檔案 ===
+    if file_path:
+        file_lower = file_path.lower()
+
+        # === 3-1: 處理 CSV ===
+        if file_lower.endswith(".csv"):
+            try:
+                with open(file_path, 'rb') as f:
+                    result = chardet.detect(f.read())
+                    encoding = result['encoding']
+                df = pd.read_csv(file_path, encoding=encoding)
+                csv_dataframe = df  # 👈 保證 global 賦值成功
+
+                # 若 query 有提到 file，加 context
+                if "file" in query.lower() or "upload" in query.lower():
+                    query = f"The user uploaded the following CSV file:\n\n{query}"
+
+                result = crew.kickoff(inputs={"query": query})
+                return safe_format_result(result)
+            except Exception as e:
+                return f"CSV Parsing Error: {e}"
+
+        # === 3-2: 處理 PDF / DOCX / TXT ===
+        elif file_lower.endswith((".pdf", ".txt", ".docx")):
+            try:
+                loader = (
+                    PyPDFLoader(file_path) if file_lower.endswith(".pdf")
+                    else UnstructuredWordDocumentLoader(file_path) if file_lower.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)
+                session_retriever = db.as_retriever()
+                session_qa_chain = ConversationalRetrievalChain.from_llm(
+                    llm=llm_gpt4,
+                    retriever=session_retriever,
+                    memory=ConversationBufferMemory(memory_key="chat_history", return_messages=True),
+                )
+
+                # 若是摘要類 Query，先用 Summarize Chain 回傳
+                if any(kw in query.lower() for kw in ["summarize", "summary", "摘要", "總結"]):
+                    return document_summarize(file_path)
+
+                # 若判斷需用 QA Chain
+                if use_file_chain:
+                    try:
+                        return session_qa_chain.run(query)
+                    except Exception as e:
+                        return f"Document QA Error: {e}"
+
+                # 否則進入 Multi-Agent 推理
+                if "file" in query.lower() or "upload" in query.lower():
+                    query = f"The user uploaded the following document:\n\n{query}"
+
+                result = crew.kickoff(inputs={"query": query})
+                return safe_format_result(result)
+
+            except Exception as e:
+                return f"Document Processing Error: {e}"
+
+        else:
+            return "Unsupported file format."
+
+    # === Step 4: 沒有上傳檔案，直接用 CrewAI 推理 ===
+    try:
+        result = crew.kickoff(inputs={"query": query})
+        return safe_format_result(result)
+    except Exception as e:
+        return f"Multi-Agent Error: {e}"
+
+# === Gradio Interface Settings ===
+demo_description = """
+**Context**:
+This demo uses a Retrieval-Augmented Generation (RAG) system based on 
+Biden’s 2023 State of the Union Address. 
+All responses are grounded in this document. 
+If no relevant information is found in the document, the system will say "No relevant info found."
+
+**Sample Questions**:
+1. What were the main topics regarding infrastructure in this speech?
+2. How does the speech address the competition with China?
+3. What does Biden say about job growth in the past two years?
+4. Does the speech mention anything about Social Security or Medicare?
+5. What does the speech propose regarding Big Tech or online privacy?
+
+*Note: The LLaMA module generates responses based solely on the current query without follow-up memory or chat history management.*
+
+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 
+Biden’s 2023 State of the Union Address. 
+All responses are grounded in this document. 
+If no relevant information is found in the document, the system will say "No relevant info found."
+
+**Sample Questions**:
+1. What were the main topics regarding infrastructure in this speech?
+2. How does the speech address the competition with China?
+3. What does Biden say about job growth in the past two years?
+4. Does the speech mention anything about Social Security or Medicare?
+5. What does the speech propose regarding Big Tech or online privacy?
+
+*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 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.
+
+Feel free to ask any question related to your document.
+"""
+demo_description4 = """
+**Context**:
+This assistant performs multi-agent tasks using tools such as:
+- Document summarisation
+- FAQ-style document Q&A
+- 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.
+"""
+demo_description5 = """
+**Context**:
+This demo uses Document Summarization via a Map-Reduce chain.
+Upload a PDF, TXT, or DOCX file to get an automatic summary of its contents.
+"""
+
+demo = gr.TabbedInterface(
+    interface_list=[
+        gr.Interface(
+            fn=rag_llama_qa,
+            inputs="text",
+            outputs="text",
+            title="Biden Q&A (LLaMA)",
+            allow_flagging="never",
+            description=demo_description
+        ),
+        gr.Interface(
+            fn=rag_gpt4_qa,
+            inputs="text",
+            outputs="text",
+            title="Biden Q&A (GPT-4)",
+            allow_flagging="never",
+            description=demo_description2
+        ),
+        gr.Interface(
+            fn=upload_and_chat,
+            inputs=[gr.File(label="Upload PDF, TXT, or DOCX"), gr.Textbox(label="Ask a question")],
+            outputs="text",
+            title="Your Docs Q&A (Upload + GPT-4)",
+            allow_flagging="never",
+            description=demo_description3
+        ),
+        gr.Interface(
+            fn=document_summarize,
+            inputs=[gr.File(label="Upload PDF, TXT, or DOCX")],
+            outputs="text",
+            title="Document Summarization",
+            allow_flagging="never",
+            description=demo_description5
+        ),
+        gr.Interface(
+            fn=multi_agent_chat_advanced,
+            inputs=[
+                gr.Textbox(label="請輸入查詢內容"),
+                gr.File(label="上傳文件 (CSV, PDF, TXT, DOCX)", file_count="single")
+            ],
+            outputs="text",
+            title="Multi-Agent AI Assistant",
+            allow_flagging="never",
+            description=demo_description4
+        )
+    ],
+    tab_names=[
+        "Biden Q&A (LLaMA)",
+        "Biden Q&A (GPT-4)",
+        "Your Docs Q&A (Upload + GPT-4)",
+        "Document Summarization",
+        "Multi-Agent AI Assistant"
+    ],
+    title="RAG + Multi-Agent Platform"
 )
 
 if __name__ == "__main__":