hi

app.py

@@ -82,7 +82,7 @@ def estimate_tokens(messages: List[Dict[str, str]]) -> int:
         total += len(msg["content"].split()) * 1.3
     return int(total)
 
-# ---------- 網路搜尋（已修復相似度 bug） ----------
+# ---------- 網路搜尋 ----------
 def perform_web_search(query: str, max_results: int = 5) -> str:
     print(f"開始網路搜尋：查詢詞 = '{query}'")
     try:
@@ -113,7 +113,7 @@ def perform_web_search(query: str, max_results: int = 5) -> str:
         
         for i, (score, result) in enumerate(relevant_with_scores[:5], 1):
             print(f"結果 {i}: 標題='{result['title']}'，相似度={score:.2f}，來源={result['url']}")
-            search_summary += f"{i}. [{score:.2f}] {result['title']}\n   {result['content'][:300]}...\n   來源: {result['url']}\n\n"
+            search_summary += f"{i}. [{score:.2f}] {result['title']}\n   {result['content'][:400]}...\n   來源: {result['url']}\n\n"
         
         return search_summary
     
@@ -135,7 +135,7 @@ class ChatPipeline:
         )
 
     @retry(stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1, min=4, max=10))
-    def _llm_call(self, messages: List[Dict[str, str]]) -> str:
+    def _llm_call(self, messages: List[Dict[str, str]], max_tokens: int = None) -> str:
         token_est = estimate_tokens(messages)
         if token_est > 50000:
             raise ValueError("輸入過長")
@@ -143,13 +143,31 @@ class ChatPipeline:
         response = self.llm_client.chat.completions.create(
             model=LLM_MODEL_CONFIG["model"],
             messages=messages,
-            max_tokens=LLM_MODEL_CONFIG["max_tokens"],
+            max_tokens=max_tokens or LLM_MODEL_CONFIG["max_tokens"],
             temperature=LLM_MODEL_CONFIG["temperature"],
             seed=LLM_MODEL_CONFIG["seed"],
             timeout=30.0,
         )
         return response.choices[0].message.content or ""
 
+    def _needs_search(self, user_text: str, history: List[Dict[str, str]]) -> bool:
+        """輕量判斷是否需要網路搜尋"""
+        router_prompt = [
+            {"role": "system", "content": "你只需要判斷用戶問題是否需要最新的網路資訊來回答。"
+                                        "如果是永恆知識（如聖經、數學原理、哲學、歷史經典等），回答 no。"
+                                        "如果是時事、新聞、最新研究、實時數據、當前事件等，回答 yes。"
+                                        "只回覆單字：yes 或 no。不要解釋。"},
+            *history,
+            {"role": "user", "content": user_text}
+        ]
+        try:
+            decision = self._llm_call(router_prompt, max_tokens=10).strip().lower()
+            print(f"搜尋需求判斷：{decision}（問題：{user_text}）")
+            return decision == "yes"
+        except Exception as e:
+            print(f"搜尋判斷失敗，預設不搜尋：{e}")
+            return False
+
     def get_conversation_history(self, user_id: str) -> List[Dict[str, str]]:
         return conversations.get(user_id, [])
 
@@ -165,23 +183,32 @@ class ChatPipeline:
             self.clear_conversation_history(user_id)
             return "對話紀錄已清除！現在開始新的對話。"
         
-        search_results = perform_web_search(user_text)
-        
         history = self.get_conversation_history(user_id)
+        
+        # ---- 新增：判斷是否需要搜尋 ----
+        needs_search = self._needs_search(user_text, history)
+        
+        search_results = None
+        if needs_search:
+            search_results = perform_web_search(user_text)
+        
+        # ---- 建構最終 prompt ----
         messages = [{"role": "system", "content": SYSTEM_PROMPT}]
         messages.extend(history)
         messages.append({"role": "user", "content": user_text})
         
-        if "沒有找到" not in search_results:
-            messages.append({"role": "system", "content": f"網路搜尋結果：{search_results}"})
+        if search_results and "沒有找到" not in search_results and "錯誤" not in search_results:
+            messages.append({"role": "system", "content": f"網路搜尋結果（僅在高度相關時使用）：{search_results}"})
         
         response = self._llm_call(messages)
         response = response.replace('*', '')
         
+        # 更新歷史（包含最終回應）
         history.append({"role": "user", "content": user_text})
         history.append({"role": "assistant", "content": response})
         self.update_conversation_history(user_id, history)
         
+        # 長回應處理
         chunks = split_text_for_line(response)
         if len(chunks) > 5:
             summary_prompt = [