Update app.py

app.py

@@ -5,7 +5,7 @@ import pandas as pd
 from typing import Optional
 from smolagents import CodeAgent, OpenAIServerModel, tool
 
-# --- 自動安裝與防卡死搜尋工具 ---
+# --- 搜尋工具設定 ---
 try:
     from duckduckgo_search import DDGS
 except ImportError:
@@ -16,21 +16,20 @@ except ImportError:
 @tool
 def web_search(query: str) -> str:
     """
-    Performs a web search.
+    Search the web for information.
     Args:
         query: The search query string.
     """
-    print(f"🕵️ [Debug] Searching for: {query}")  # 讓你知道它正在工作
+    print(f"🕵️ [Debug] Searching: {query}")
     try:
-        # 關鍵修改：加上 timeout=15 (秒)，並限制結果數量以加速
-        # backend='html' 通常比 api 模式更不容易被擋
-        results = DDGS(timeout=15).text(query, max_results=3)
+        # 設定 10 秒逾時，只抓 2 筆結果，降低負擔
+        results = DDGS(timeout=10).text(query, max_results=2)
         if not results:
-            return "No search results found."
+            return "No results."
         return str(results)
     except Exception as e:
-        print(f"❌ [Error] Search failed: {e}")
-        return f"Search connection failed: {e}"
+        print(f"❌ Search Error: {e}")
+        return "Search failed."
 
 # -----------------------------------------------------------
 
@@ -43,10 +42,9 @@ class GroqAgent:
             self.agent = None
             return
             
-        # 原本是 llama-3.3-70b-versatile
-        # 請改成下面這個：
+        # 修正：換回目前 Groq 支援的主力模型
         model = OpenAIServerModel(
-            model_id="mixtral-8x7b-32768",  # <--- 改用這個速度最快的模型
+            model_id="llama-3.3-70b-versatile",
             api_base="https://api.groq.com/openai/v1",
             api_key=self.api_key
         )
@@ -54,7 +52,7 @@ class GroqAgent:
         self.agent = CodeAgent(
             tools=[web_search], 
             model=model,
-            max_steps=4,
+            max_steps=3, # 減少步數，避免想太久卡住
             verbosity_level=1
         )
     
@@ -63,12 +61,11 @@ class GroqAgent:
             return "Error: GROQ_API_KEY not configured."
         
         try:
-            # 簡化 Prompt 讓模型反應更快
+            # 極簡化 Prompt，讓模型反應更快
             prompt = f"""
-            You are an expert agent. Answer the question concisely.
-            1. Use 'web_search' for facts (dates, names, events).
-            2. If search fails, make your best guess.
-            3. Answer directly.
+            Task: Answer the question concisely.
+            Tool: Use 'web_search' only for specific facts (names, dates).
+            Fallback: If search fails, guess.
             
             Question: {question}
             """
@@ -106,16 +103,16 @@ def run_and_submit_all(profile: Optional[gr.OAuthProfile] = None):
         q = item.get("question")
         tid = item.get("task_id")
         
-        print(f"🚀 [{idx}/{total}] Processing task: {tid}...")
+        print(f"🚀 [{idx}/{total}] Task: {tid}")
         
-        # 這裡會觸發 agent 思考
+        # 執行 Agent
         ans = agent_wrapper(q)
         
         answers.append({"task_id": tid, "submitted_answer": ans})
-        logs.append({"Task": tid, "Status": "Done", "Answer Preview": ans[:50]})
+        logs.append({"Task": tid, "Answer": ans[:50] + "..."})
 
     try:
-        print("Submitting answers...")
+        print("Submitting...")
         res = requests.post(f"{api_url}/submit", json={
             "username": username,
             "agent_code": f"https://huggingface.co/spaces/{space_id}/tree/main",
@@ -124,20 +121,21 @@ def run_and_submit_all(profile: Optional[gr.OAuthProfile] = None):
         
         data = res.json()
         score = data.get('score', 0)
-        msg = f"🎉 Final Score: {score}%" if score >= 30 else f"Score: {score}% (Try again!)"
+        
+        # 只要有一題對就是成功，通常文字搜尋題會對，圖片題會錯
+        msg = f"🎉 Final Score: {score}%" 
         return msg, pd.DataFrame(logs)
         
     except Exception as e:
         return f"Submit error: {str(e)}", pd.DataFrame(logs)
 
-with gr.Blocks(title="Final Agent (Timeout Optimized)") as demo:
-    gr.Markdown("# 🚀 Final Agent (Fast Version)")
-    gr.Markdown("Has timeout protection to prevent hanging.")
+with gr.Blocks(title="Final Agent (Fixed Model)") as demo:
+    gr.Markdown("# 🚀 Final Agent (Llama 3.3)")
     with gr.Row():
         gr.LoginButton()
         btn = gr.Button("Run Evaluation", variant="primary")
     out = gr.Textbox(label="Status")
-    tab = gr.DataFrame(label="Progress Log")
+    tab = gr.DataFrame(label="Logs")
     btn.click(run_and_submit_all, outputs=[out, tab])
 
 if __name__ == "__main__":