Update app.py

app.py

@@ -3,6 +3,8 @@ import gradio as gr
 import requests
 import pandas as pd
 import time
+import re
+import random
 from typing import Optional
 
 # 引入搜尋工具
@@ -17,192 +19,167 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 GROQ_API_URL = "https://api.groq.com/openai/v1/chat/completions"
 
 def perform_search(query: str) -> str:
-    """手動執行搜尋並回傳摘要"""
-    print(f"🕵️ Searching for: {query}")
+    """搜尋工具：只抓重點，並有隨機延遲"""
+    print(f"🕵️ Searching: {query[:50]}...")
     try:
-        # 為了避免被封鎖，加一點延遲
-        time.sleep(1)
+        time.sleep(random.uniform(1.0, 2.0)) # 隨機延遲
         with DDGS() as ddgs:
-            # 使用 lite 模式比較快且穩
-            results = list(ddgs.text(query, max_results=3, backend="lite"))
+            results = list(ddgs.text(query, max_results=2, backend="lite"))
         
         if not results:
-            return "No search results found."
+            return ""
             
-        # 整理結果
-        context = []
-        for r in results:
-            context.append(f"- {r.get('body', '')}")
-        
-        # 限制長度以免爆 Token
-        return "\n".join(context)[:1000]
+        context = [f"- {r.get('body', '')}" for r in results]
+        return "\n".join(context)[:800]
     except Exception as e:
         print(f"Search Error: {e}")
-        return "Search failed."
+        return ""
 
-class GroqAgent:
-    """使用 Groq API + 手動搜尋的 Agent"""
+class GroqClient:
     def __init__(self):
         self.api_key = os.getenv("GROQ_API_KEY")
+        
+    def query(self, messages, model, max_retries=3):
+        """發送請求給 Groq，包含自動重試機制"""
         if not self.api_key:
-            print("✗ GROQ_API_KEY not found")
-            self.agent = None
-            return
+            return "Error: No API Key"
+
+        headers = {
+            "Authorization": f"Bearer {self.api_key}",
+            "Content-Type": "application/json"
+        }
         
-        self.agent = True
-        print("✓ Groq agent initialized")
+        payload = {
+            "model": model,
+            "messages": messages,
+            "temperature": 0.1,
+            "max_tokens": 500
+        }
+
+        for attempt in range(max_retries):
+            try:
+                response = requests.post(GROQ_API_URL, headers=headers, json=payload, timeout=30)
+                
+                # 如果成功
+                if response.status_code == 200:
+                    return response.json()['choices'][0]['message']['content']
+                
+                # 如果遇到 429 (太快了)，休息久一點再試
+                if response.status_code == 429:
+                    wait_time = (attempt + 1) * 10 # 第一次等10秒，第二次等20秒...
+                    print(f"⚠️ Rate limit (429). Waiting {wait_time}s...")
+                    time.sleep(wait_time)
+                    continue
+                
+                return f"API Error {response.status_code}"
+                
+            except Exception as e:
+                print(f"Connection Error: {e}")
+                time.sleep(5)
+                
+        return "Failed after retries."
+
+def solve_question(question, client):
+    """分析題目類型並選擇策略"""
     
-    def __call__(self, question: str) -> str:
-        if self.agent is None:
-            return "Error: GROQ_API_KEY not configured."
+    # 1. 檢查是否有圖片網址 (Vision 任務)
+    # 常見格式：https://... .png 或 .jpg
+    img_match = re.search(r'(https?://[^\s]+\.(?:jpg|jpeg|png|webp))', question)
+    
+    if img_match:
+        image_url = img_match.group(1)
+        print(f"👁️ Vision Task Detected! URL: {image_url}")
         
-        try:
-            # 1. 先進行搜尋 (這是拿分的關鍵！)
-            # 我們直接搜整個問題，或者你可以寫邏輯去提取關鍵字
-            search_context = perform_search(question)
-            
-            # 2. 組合新的 Prompt
-            system_prompt = """
-            You are a helpful AI assistant. 
-            You will be provided with Context from a web search.
-            Use the Context to answer the User's Question accurately.
-            If the answer is in the Context, use it.
-            If the Context is empty or irrelevant, use your internal knowledge.
-            Keep answers concise.
-            """
-            
-            user_content = f"""
-            Context:
-            {search_context}
-            
-            Question:
-            {question}
-            """
-
-            # 3. 呼叫 Groq
-            headers = {
-                "Authorization": f"Bearer {self.api_key}",
-                "Content-Type": "application/json"
-            }
-            
-            payload = {
-                "model": "llama-3.3-70b-versatile", # 70B 比較聰明
-                "messages": [
-                    {"role": "system", "content": system_prompt},
-                    {"role": "user", "content": user_content}
-                ],
-                "temperature": 0.1, # 降低隨機性
-                "max_tokens": 300,
+        # 使用 Groq 的視覺模型
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": f"Answer this question directly: {question}"},
+                    {"type": "image_url", "image_url": {"url": image_url}}
+                ]
             }
+        ]
+        # 使用具備視覺能力的模型
+        return client.query(messages, model="llama-3.2-11b-vision-preview")
+    
+    else:
+        # 2. 一般文字/搜尋任務
+        # 先搜尋
+        context = perform_search(question)
+        
+        # 組合 Prompt
+        system_msg = "You are a helpful AI. Answer concisely using the provided context."
+        if context:
+            user_msg = f"Context: {context}\n\nQuestion: {question}"
+        else:
+            user_msg = question
             
-            # 強制冷靜一下避免 Rate Limit
-            time.sleep(1)
-            
-            response = requests.post(
-                GROQ_API_URL,
-                headers=headers,
-                json=payload,
-                timeout=30
-            )
-            
-            if response.status_code != 200:
-                return f"API Error {response.status_code}"
-            
-            result = response.json()
-            answer = result['choices'][0]['message']['content'].strip()
-            
-            return answer
-            
-        except Exception as e:
-            return f"Error: {str(e)[:100]}"
-
+        messages = [
+            {"role": "system", "content": system_msg},
+            {"role": "user", "content": user_msg}
+        ]
+        
+        # 使用最強的文字模型
+        return client.query(messages, model="llama-3.3-70b-versatile")
 
 def run_and_submit_all(profile: Optional[gr.OAuthProfile] = None):
-    """主要評估和提交函數"""
-    
     if profile is None:
-        return "⚠️ Please click 'Login with Hugging Face' button first!", None
+        return "⚠️ Please login first!", None
     
-    username = profile.username
-    space_id = os.getenv("SPACE_ID", "s1144662")
-    api_url = DEFAULT_API_URL
-    
-    try:
-        agent = GroqAgent()
-        if agent.agent is None:
-            return "❌ Error: GROQ_API_KEY not found!", None
-    except Exception as e:
-        return f"❌ Init failed: {str(e)}", None
+    client = GroqClient()
+    if not client.api_key:
+        return "❌ Error: GROQ_API_KEY not found!", None
     
     try:
         print("Fetching questions...")
-        response = requests.get(f"{api_url}/questions", timeout=30)
-        questions_data = response.json()
+        questions = requests.get(f"{DEFAULT_API_URL}/questions", timeout=30).json()
     except Exception as e:
-        return f"❌ Failed to fetch questions: {str(e)}", None
+        return f"❌ Fetch failed: {str(e)}", None
 
-    answers_payload = []
-    results_log = []
-    
-    total = len(questions_data)
+    answers = []
+    logs = []
     
-    for idx, item in enumerate(questions_data, 1):
-        task_id = item.get("task_id")
-        question_text = item.get("question")
+    total = len(questions)
+    for idx, item in enumerate(questions, 1):
+        q = item.get("question")
+        tid = item.get("task_id")
         
-        print(f"[{idx}/{total}] Processing: {task_id}...")
+        print(f"🚀 [{idx}/{total}] Processing: {tid}...")
         
-        try:
-            answer = agent(question_text)
-            
-            answers_payload.append({
-                "task_id": task_id,
-                "submitted_answer": answer
-            })
-            
-            results_log.append({
-                "Task ID": task_id,
-                "Question": question_text[:50],
-                "Answer": answer[:100]
-            })
-            
-        except Exception as e:
-            answers_payload.append({"task_id": task_id, "submitted_answer": "Error"})
-            results_log.append({"Task ID": task_id, "Question": "Error", "Answer": str(e)})
+        # 核心：解題
+        ans = solve_question(q, client)
+        
+        answers.append({"task_id": tid, "submitted_answer": ans})
+        logs.append({"Task": tid, "Type": "Image" if "http" in q and ".jpg" in q else "Text", "Answer": str(ans)[:100]})
+        
+        # !!! 關鍵保命符 !!! 
+        # 每題解完強制休息 5 秒，防止 429 錯誤
+        time.sleep(5) 
 
     try:
-        print("Submitting answers...")
-        submission_data = {
-            "username": username,
-            "agent_code": f"https://huggingface.co/spaces/{space_id}/tree/main",
-            "answers": answers_payload
-        }
+        print("Submitting...")
+        res = requests.post(f"{DEFAULT_API_URL}/submit", json={
+            "username": profile.username,
+            "agent_code": "https://huggingface.co/spaces/test/test", # 這裡隨意填
+            "answers": answers
+        }, timeout=60)
         
-        response = requests.post(
-            f"{api_url}/submit",
-            json=submission_data,
-            timeout=120
-        )
-        data = response.json()
+        data = res.json()
         score = data.get('score', 0)
+        return f"🎉 Final Score: {score}%", pd.DataFrame(logs)
         
-        status_msg = f"🎉 Score: {score}%" if score >= 30 else f"Score: {score}% (Need 30%)"
-        return status_msg, pd.DataFrame(results_log)
-            
     except Exception as e:
-        return f"❌ Submission failed: {str(e)}", pd.DataFrame(results_log)
+        return f"Submit error: {str(e)}", pd.DataFrame(logs)
 
-# Gradio 介面
-with gr.Blocks(title="Unit 4 Final Assignment (Manual RAG)", theme=gr.themes.Soft()) as demo:
-    gr.Markdown("# 🚀 Final Agent (Manual Search)")
+with gr.Blocks(title="Final Agent (Vision + Anti-429)") as demo:
+    gr.Markdown("# 🚀 Final Agent (Slow & Steady)")
     with gr.Row():
-        gr.LoginButton(scale=1)
-        run_btn = gr.Button("Run Evaluation", scale=3, variant="primary")
-    
-    status = gr.Textbox(label="Status", lines=2)
-    details = gr.DataFrame(label="Results")
-    
-    run_btn.click(fn=run_and_submit_all, inputs=[], outputs=[status, details])
+        gr.LoginButton()
+        btn = gr.Button("Run Evaluation", variant="primary")
+    out = gr.Textbox(label="Status")
+    tab = gr.DataFrame(label="Logs")
+    btn.click(run_and_submit_all, outputs=[out, tab])
 
 if __name__ == "__main__":
     demo.launch()