Update app.py

app.py

@@ -1238,15 +1238,56 @@ def multi_agent_chat_advanced(query: str, file=None) -> str:
 
 
 # LangGraph 使用的節點函數（會接續你的 Crew Agent）
+from sentence_transformers import SentenceTransformer
+
+embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
+
+# Intent label embeddings（你可以再自訂）
+INTENT_LABELS = {
+    "DocQA": ["document", "file", "paper", "cb", "proposal"],
+    "Summarise": ["summarise", "summary", "abstract", "key points", "要點", "總結", "摘要"],
+    "General": ["who are you", "tell me something", "what can you do", "fun fact"],
+}
+
+def detect_intent_embedding(query, file_names=[]):
+    query_emb = embedding_model.encode(query, normalize_embeddings=True)
+
+    best_label = None
+    best_score = -1
+
+    # 把上傳的檔名也當成一種 prompt
+    all_phrases = INTENT_LABELS.copy()
+    if file_names:
+        all_phrases["DocQA"] += [name.lower() for name in file_names]
+
+    for label, examples in all_phrases.items():
+        for example in examples:
+            example_emb = embedding_model.encode(example, normalize_embeddings=True)
+            score = float(query_emb @ example_emb.T)
+            if score > best_score:
+                best_score = score
+                best_label = label
+    return best_label if best_label else "General"
+
+def decide_next(state):
+    query = state.get("query", "")
+    file_names = state.get("file_names", [])
+    label = detect_intent_embedding(query, file_names)
+    return label
+
+# General → Search fallback handoff（當 general 回應不了）
+def general_run(state: dict) -> dict:
+    result = general_agent.kickoff(inputs={"query": state["query"]})
+    output = result.output.lower()
+    if any(x in output for x in ["i don't know", "no idea", "not sure", "can't answer", "no info"]):
+        result = search_agent.kickoff(inputs={"query": state["query"]})
+    return {"query": state["query"], "final_answer": result.output}
+    
 def router_run(state: dict) -> dict:
     query = state["query"]
     result = router_agent.kickoff(inputs={"query": query})
     return {"query": query, "route_result": result.output}
 
-def general_run(state: dict) -> dict:
-    result = general_agent.kickoff(inputs={"query": state["query"]})
-    return {"query": state["query"], "final_answer": result.output}
-
 def docqa_run(state: dict) -> dict:
     result = document_qa_agent.kickoff(inputs={"query": state["query"]})
     return {"query": state["query"], "final_answer": result.output}
@@ -1268,13 +1309,10 @@ def build_langgraph_gpt_like():
 
     # Router 條件邏輯：你可以換成你自己的邏輯規則
     def decide_next(state):
-        route = state.get("route_result", "").lower()
-        if "summarise" in route or "summary" in route:
-            return "Summarise"
-        elif "document" in route or "file" in route:
-            return "DocQA"
-        else:
-            return "General"
+        query = state.get("query", "")
+        file_names = state.get("file_names", [])
+        label = detect_intent_embedding(query, file_names)
+        return label
 
     graph.add_conditional_edges("Router", decide_next, {
         "DocQA": "DocQA",
@@ -1289,28 +1327,43 @@ def build_langgraph_gpt_like():
     return graph.compile()
 
 # 用於 Gradio Tab 6 的主入口函數
-def langgraph_tab6_main(query: str, file=None) -> str:
+def langgraph_tab6_main(query: str, files=None) -> str:
     try:
-        file_path = get_file_path(file)
-        if not file_path:
-            return "Please upload a document file first."
-
-        # 檢查文件類型並建立 retriever
-        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)
+        if not files:
+            return "Please upload at least one document file."
+
+        all_docs = []
+        file_names = []
+
+        for file in files:
+            file_path = get_file_path(file)
+            if not file_path:
+                continue
+            file_names.append(os.path.basename(file_path))
+            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()
+            all_docs.extend(docs)
+
+        if not all_docs:
+            return "No valid documents loaded."
+
+        chunks = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50).split_documents(all_docs)
         db = FAISS.from_documents(chunks, embeddings)
         retriever = db.as_retriever()
 
-        # 建構 LangGraph vB QA Chain
+        # 建構 LangGraph QA Chain
         graph = build_langgraph_doc_qa_chain(llm=llm_gpt4, retriever=retriever, memory=None, prompt=custom_prompt)
-        result = graph(query)
+        result = graph.invoke({
+            "query": query,
+            "file_names": file_names  # Pass all file names for intent detection
+        })
         return result.get("answer", "No answer generated.")
+
     except Exception as e:
         return f"[Tab6 Error] {e}"
     
@@ -1459,7 +1512,10 @@ demo = gr.TabbedInterface(
         ),
         gr.Interface(
             fn=langgraph_tab6_main,
-            inputs=[gr.Textbox(label="Ask anything")],
+            inputs=[
+                gr.Textbox(label="Ask anything"),
+                gr.File(label="Upload one or more files", file_types=[".pdf", ".txt", ".docx"], file_count="multiple")
+            ],
             outputs="text",
             title="LangGraph GPT-like QA (Tab6)",
             allow_flagging="never",