Update app.py

app.py

@@ -27,7 +27,6 @@ from youtube_transcript_api import YouTubeTranscriptApi
 # 配置常量
 # =============================================================================
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
-# AGICTO_BASE_URL 请设置为 https://api.agicto.cn （不含 /v1）
 AGICTO_BASE_URL = os.getenv("AGICTO_BASE_URL", "https://api.agicto.cn")
 AGICTO_API_KEY = os.getenv("AGICTO_API_KEY", "")
 QWEN_MODEL = "qwen3.5-35b-a3b"
@@ -36,7 +35,6 @@ QWEN_MODEL = "qwen3.5-35b-a3b"
 # 进度监控器
 # =============================================================================
 class ProgressMonitor:
-    # ... 保持不变 ...
     def __init__(self):
         self.current = 0
         self.total = 0
@@ -80,16 +78,15 @@ class ProgressMonitor:
             return html
 
 # =============================================================================
-# Qwen LLM 封装（修正 URL 拼接）
+# Qwen LLM 封装
 # =============================================================================
 class QwenLLM:
     def __init__(self, model=QWEN_MODEL):
         self.model = model
         self.api_key = AGICTO_API_KEY
-        # 规范化 base_url，确保末尾没有多余斜杠，并去掉可能存在的 /v1
         base = AGICTO_BASE_URL.rstrip('/')
         if base.endswith('/v1'):
-            base = base[:-3]  # 去掉 /v1
+            base = base[:-3]
         self.base_url = base
         if not self.api_key:
             print("⚠️ 未设置 AGICTO_API_KEY，请检查环境变量")
@@ -108,8 +105,6 @@ class QwenLLM:
         if functions:
             body["tools"] = [{"type": "function", "function": f} for f in functions]
             body["tool_choice"] = "auto"
-
-        # 统一使用 /v1/chat/completions 路径
         url = f"{self.base_url}/v1/chat/completions"
         try:
             resp = requests.post(url, headers=headers, json=body, timeout=60)
@@ -190,7 +185,7 @@ class QwenLLM:
         return formatted
 
 # =============================================================================
-# 工具定义（analyze_image / transcribe_audio 也需要使用修正后的 URL）
+# 工具定义（所有工具均附带 description）
 # =============================================================================
 api_url_tasks = DEFAULT_API_URL
 
@@ -322,16 +317,16 @@ def download_file_for_task(task_id: str) -> str:
         return f"文件下载失败: {e}"
 
 # =============================================================================
-# LangGraph Agent
+# LangGraph 状态与节点
 # =============================================================================
 class AgentState(TypedDict):
     messages: Annotated[Sequence[BaseMessage], operator.add]
     final_answer: str
     task_id: str
+    tool_attempts: int   # 已执行工具调用次数
 
 tools = [web_search, web_scraper, calculator, analyze_image, transcribe_audio, get_youtube_transcript, download_file_for_task]
 tool_node = ToolNode(tools)
-
 llm = QwenLLM()
 functions = [convert_to_openai_function(t) for t in tools]
 llm_with_tools = llm.bind_functions(functions)
@@ -339,22 +334,50 @@ llm_with_tools = llm.bind_functions(functions)
 def agent_node(state: AgentState) -> dict:
     messages = state["messages"]
     task_id = state.get("task_id", "")
-    sys_prompt = f"""You are a helpful assistant answering GAIA Level 1 questions. Use tools if needed. 
-When you know the answer, output only the answer string, without any extra text or "FINAL ANSWER:".
-Current task ID: {task_id}. If you need the file for this task, use download_file_for_task with task_id="{task_id}"."""
+    sys_prompt = f"""You are a helpful assistant answering GAIA Level 1 questions. 
+IMPORTANT: You MUST use at least one tool (e.g., web_search, web_scraper, download_file_for_task) to verify or retrieve information, even if you think you already know the answer.
+When you have the final answer, output only the answer string, without any extra text or "FINAL ANSWER:".
+Current task ID: {task_id}. If the question requires a file, use download_file_for_task with task_id="{task_id}"."""
     full = [SystemMessage(content=sys_prompt)] + list(messages)
     response = llm_with_tools.invoke(full)
     return {"messages": [response]}
 
 def should_continue(state: AgentState) -> str:
-    last = state["messages"][-1]
+    messages = state["messages"]
+    last = messages[-1]
+    tool_attempts = state.get("tool_attempts", 0)
+    MAX_TOOL_CALLS = 5
+
+    # 超过最大调用次数，强制结束
+    if tool_attempts >= MAX_TOOL_CALLS:
+        return "finish"
+
+    # 如果 LLM 请求了工具调用，允许执行
     if hasattr(last, "additional_kwargs") and "function_call" in last.additional_kwargs:
         return "tools"
+
+    # 尚未调用过任何工具？强制要求使用工具
+    tool_msg_count = sum(1 for m in messages if isinstance(m, ToolMessage))
+    if tool_msg_count == 0:
+        return "force_tool"
+
+    # 已经用过工具，可以结束
     return "finish"
 
+def force_tool_node(state: AgentState) -> dict:
+    new_msg = HumanMessage(
+        content="You have not used any tools yet. Please use at least one tool to find or verify the answer. "
+                "Search the web, download a file, or analyze an image if provided."
+    )
+    return {"messages": [new_msg]}
+
+def increment_tool_count(state: AgentState) -> dict:
+    return {"tool_attempts": state.get("tool_attempts", 0) + 1}
+
 def finish_node(state: AgentState) -> dict:
     last = state["messages"][-1]
     content = last.content
+    # 提取最终答案（纯文本）
     answer = content.strip().split("\n")[-1].strip()
     if "FINAL ANSWER:" in answer:
         answer = answer.split("FINAL ANSWER:")[-1].strip()
@@ -365,19 +388,37 @@ def build_graph():
     workflow.add_node("agent", agent_node)
     workflow.add_node("tools", tool_node)
     workflow.add_node("finish", finish_node)
+    workflow.add_node("force_tool", force_tool_node)
+    workflow.add_node("count_tools", increment_tool_count)
+
     workflow.set_entry_point("agent")
-    workflow.add_conditional_edges("agent", should_continue, {"tools": "tools", "finish": "finish"})
-    workflow.add_edge("tools", "agent")
+    workflow.add_conditional_edges(
+        "agent",
+        should_continue,
+        {"tools": "tools", "force_tool": "force_tool", "finish": "finish"}
+    )
+    workflow.add_edge("tools", "count_tools")
+    workflow.add_edge("count_tools", "agent")
+    workflow.add_edge("force_tool", "agent")
     workflow.add_edge("finish", END)
+
     return workflow.compile()
 
+# =============================================================================
+# Agent 类
+# =============================================================================
 class LangGraphAgent:
     def __init__(self):
         self.graph = build_graph()
         print("LangGraphAgent 初始化完成，使用模型:", QWEN_MODEL)
 
     def __call__(self, question: str, task_id: str = "") -> str:
-        state = {"messages": [HumanMessage(content=question)], "final_answer": "", "task_id": task_id}
+        state = {
+            "messages": [HumanMessage(content=question)],
+            "final_answer": "",
+            "task_id": task_id,
+            "tool_attempts": 0
+        }
         try:
             final_state = self.graph.invoke(state)
             return final_state["final_answer"]
@@ -386,7 +427,7 @@ class LangGraphAgent:
             return f"Error: {e}"
 
 # =============================================================================
-# 主运行函数
+# 主运行函数（生成器，实时进度）
 # =============================================================================
 import pandas as pd
 
@@ -479,7 +520,6 @@ with gr.Blocks(title="GAIA Agent") as demo:
     )
 
 if __name__ == "__main__":
-    # 检查必要环境变量
     if not AGICTO_API_KEY:
         print("❌ 错误：AGICTO_API_KEY 未设置！请在 Space 的 Settings -> Repository Secrets 中添加。")
     if "v1" in AGICTO_BASE_URL: