Update app.py

app.py

@@ -44,15 +44,32 @@ client = InferenceClient(api_key=hf_token) if hf_token else InferenceClient()
 print(f"✅ LLM initialized: Qwen/Qwen2.5-72B-Instruct:novita")
 print(f"📊 MCP Services: {len(MCP_SERVICES)} services, {len(MCP_TOOLS)} tools")
 
-# ========== 系统提示词(简化) ==========
-from datetime import datetime
+# ========== Token 限制配置 ==========
+# HuggingFace Inference API 实际限制约 8000-16000 tokens
+# 为了安全，设置更低的限制
+MAX_TOTAL_TOKENS = 6000  # 总上下文限制
+MAX_TOOL_RESULT_CHARS = 800  # 工具返回最大字符数 (约400 tokens)
+MAX_HISTORY_CHARS = 500  # 单条历史消息最大字符数
+MAX_HISTORY_TURNS = 2  # 最大历史轮数
+MAX_TOOL_ITERATIONS = 2  # 最大工具调用轮数
+MAX_OUTPUT_TOKENS = 1000  # 最大输出 tokens
+
+def estimate_tokens(text):
+    """估算文本 token 数量（粗略：1 token ≈ 2 字符）"""
+    return len(str(text)) // 2
+
+def truncate_text(text, max_chars, suffix="...[truncated]"):
+    """截断文本到指定长度"""
+    text = str(text)
+    if len(text) <= max_chars:
+        return text
+    return text[:max_chars] + suffix
 
 def get_system_prompt():
-    """生成包含当前日期的系统提示词"""
+    """生成包含当前日期的系统提示词（精简版）"""
+    from datetime import datetime
     current_date = datetime.now().strftime("%Y-%m-%d")
-    return f"""You are a financial analysis assistant. Use tools to get data on company financials (past 5-year reports), current stock prices, market news, and company news. Provide data-driven insights.
-
-IMPORTANT: Today's date is {current_date}. When querying news or time-sensitive data, use recent dates relative to today."""
+    return f"""Financial analyst. Today: {current_date}. Use tools for company data, stock prices, news. Be concise."""
 
 # ============================================================
 # MCP 服务调用核心代码区
@@ -159,28 +176,30 @@ def chatbot_response(message, history):
     try:
         messages = [{"role": "system", "content": get_system_prompt()}]
         
-        # 添加历史(最近3轭) - 减少上下文长度
+        # 添加历史(最近2轮) - 严格限制上下文长度
         if history:
-            for item in history[-3:]:  # 从5轮改为3轮
+            for item in history[-MAX_HISTORY_TURNS:]:
                 if isinstance(item, (list, tuple)) and len(item) == 2:
+                    # 用户消息（不截断）
                     messages.append({"role": "user", "content": item[0]})
-                    # 截断过长的历史回复
-                    assistant_msg = item[1]
-                    if len(assistant_msg) > 1000:
-                        assistant_msg = assistant_msg[:1000] + "...[truncated]"
+                    
+                    # 助手回复（严格截断）
+                    assistant_msg = str(item[1])
+                    if len(assistant_msg) > MAX_HISTORY_CHARS:
+                        assistant_msg = truncate_text(assistant_msg, MAX_HISTORY_CHARS)
                     messages.append({"role": "assistant", "content": assistant_msg})
         
         messages.append({"role": "user", "content": message})
         
         tool_calls_log = []
         
-        # LLM 调用循环(最多3轮工具调用) - 减少迭代次数
-        for iteration in range(3):  # 从5轮改为3轮
+        # LLM 调用循环(最多2轮工具调用) - 严格控制迭代次数
+        for iteration in range(MAX_TOOL_ITERATIONS):
             response = client.chat_completion(
                 messages=messages,
                 model="Qwen/Qwen2.5-72B-Instruct:novita",
                 tools=MCP_TOOLS,
-                max_tokens=1500,  # 从2000降到1500
+                max_tokens=MAX_OUTPUT_TOKENS,
                 temperature=0.5,
                 tool_choice="auto",
                 stream=False
@@ -198,16 +217,29 @@ def chatbot_response(message, history):
                     # 调用 MCP 工具
                     tool_result = call_mcp_tool(tool_name, tool_args)
                     
-                    # 大幅限制返回结果大小，避免超长内容导致500错误
+                    # 严格限制返回结果大小，避免超长内容导致500错误
                     result_str = json.dumps(tool_result, ensure_ascii=False)
-                    if len(result_str) > 2000:  # 从4000降到2000
-                        # 截断过长的结果，只保留关键信息
+                    
+                    # 截断到安全长度 (800字符 ≈ 400 tokens)
+                    if len(result_str) > MAX_TOOL_RESULT_CHARS:
                         if isinstance(tool_result, dict) and "text" in tool_result:
-                            # 如果是文本格式，截取前1500字符
-                            tool_result_truncated = {"text": tool_result["text"][:1500] + "...[truncated]", "_truncated": True}
+                            # 如果是文本格式
+                            truncated_text = truncate_text(tool_result["text"], MAX_TOOL_RESULT_CHARS - 50)
+                            tool_result_truncated = {"text": truncated_text, "_truncated": True}
+                        elif isinstance(tool_result, dict):
+                            # JSON 格式，保留关键字段
+                            truncated = {}
+                            char_count = 0
+                            for k, v in list(tool_result.items())[:5]:  # 只保留前5个字段
+                                v_str = str(v)[:200]  # 每个值最多200字符
+                                truncated[k] = v_str
+                                char_count += len(k) + len(v_str)
+                                if char_count > MAX_TOOL_RESULT_CHARS:
+                                    break
+                            tool_result_truncated = {**truncated, "_truncated": True}
                         else:
-                            tool_result_truncated = {"_truncated": True, "preview": result_str[:1500] + "...[truncated]"}
-                        result_for_llm = json.dumps(tool_result_truncated)
+                            tool_result_truncated = {"preview": truncate_text(result_str, MAX_TOOL_RESULT_CHARS), "_truncated": True}
+                        result_for_llm = json.dumps(tool_result_truncated, ensure_ascii=False)
                     else:
                         result_for_llm = result_str
                     
@@ -230,6 +262,27 @@ def chatbot_response(message, history):
         
         # 显示工具调用（带展开/折叠按钮）
         if tool_calls_log:
+            # 先添加 JavaScript 函数（确保在 HTML 之前加载）
+            response_prefix += """<script>
+if (typeof window.toggleToolResult === 'undefined') {
+  window.toggleToolResult = function(id) {
+    var content = document.getElementById('content_' + id);
+    var arrow = document.getElementById('arrow_' + id);
+    if (content && arrow) {
+      if (content.style.display === 'none' || content.style.display === '') {
+        content.style.display = 'block';
+        arrow.textContent = '▼';
+      } else {
+        content.style.display = 'none';
+        arrow.textContent = '▶';
+      }
+    }
+  };
+}
+</script>
+
+"""
+            
             response_prefix += """<div style='margin-bottom: 15px;'>
 <div style='background: #f0f0f0; padding: 8px 12px; border-radius: 6px; font-weight: 600; color: #333;'>
 🛠️ Tools Used ({} calls)
@@ -237,41 +290,29 @@ def chatbot_response(message, history):
 """.format(len(tool_calls_log))
             
             for idx, tool_call in enumerate(tool_calls_log):
-                tool_id = f"tool_{idx}_{hash(str(tool_call))}"
+                tool_id = f"tool{idx}"
                 
                 # 工具卡片
                 response_prefix += f"""<div style='margin: 8px 0; border: 1px solid #ddd; border-radius: 6px; overflow: hidden;'>
-  <div style='background: #fff; padding: 10px; cursor: pointer; display: flex; justify-content: space-between; align-items: center;' onclick='toggleTool("{tool_id}")'>
-    <div>
-      <strong style='color: #2c5aa0;'>📌 {idx+1}. {tool_call['name']}</strong>
-      <div style='font-size: 0.85em; color: #666; margin-top: 4px;'>📥 Input: <code>{json.dumps(tool_call['arguments'], ensure_ascii=False)}</code></div>
+  <div style='background: #fff; padding: 10px; cursor: pointer; user-select: none;' onclick='window.toggleToolResult("{tool_id}")'>
+    <div style='display: flex; justify-content: space-between; align-items: center;'>
+      <div style='flex: 1;'>
+        <strong style='color: #2c5aa0;'>📌 {idx+1}. {tool_call['name']}</strong>
+        <div style='font-size: 0.85em; color: #666; margin-top: 4px;'>📥 Input: <code style='background: #f5f5f5; padding: 2px 6px; border-radius: 3px;'>{json.dumps(tool_call['arguments'], ensure_ascii=False)}</code></div>
+      </div>
+      <span id='arrow_{tool_id}' style='font-size: 1.2em; color: #999; margin-left: 10px;'>▶</span>
     </div>
-    <span id='arrow_{tool_id}' style='font-size: 1.2em; color: #999;'>▶</span>
   </div>
-  <div id='{tool_id}' style='display: none; background: #f9f9f9; padding: 12px; border-top: 1px solid #eee;'>
+  <div id='content_{tool_id}' style='display: none; background: #f9f9f9; padding: 12px; border-top: 1px solid #eee;'>
     <div style='font-size: 0.9em; color: #333;'>
       <strong>📤 Output:</strong>
-      <pre style='background: #fff; padding: 10px; border-radius: 4px; overflow-x: auto; margin-top: 6px; font-size: 0.85em; border: 1px solid #e0e0e0;'>{json.dumps(tool_call.get('result', {}), ensure_ascii=False, indent=2)[:1000]}{'...' if len(json.dumps(tool_call.get('result', {}), ensure_ascii=False)) > 1000 else ''}</pre>
+      <pre style='background: #fff; padding: 10px; border-radius: 4px; overflow-x: auto; margin-top: 6px; font-size: 0.85em; border: 1px solid #e0e0e0; max-height: 400px;'>{json.dumps(tool_call.get('result', {}), ensure_ascii=False, indent=2)[:1500]}{'...' if len(json.dumps(tool_call.get('result', {}), ensure_ascii=False)) > 1500 else ''}</pre>
     </div>
   </div>
 </div>
 """
             
-            # JavaScript 函数
-            response_prefix += """<script>
-function toggleTool(id) {
-  var content = document.getElementById(id);
-  var arrow = document.getElementById('arrow_' + id);
-  if (content.style.display === 'none') {
-    content.style.display = 'block';
-    arrow.innerHTML = '▼';
-  } else {
-    content.style.display = 'none';
-    arrow.innerHTML = '▶';
-  }
-}
-</script>
-</div>
+            response_prefix += """</div>
 
 ---
 
@@ -285,7 +326,7 @@ function toggleTool(id) {
             messages=messages,
             model="Qwen/Qwen2.5-72B-Instruct:novita",
             tools=MCP_TOOLS,
-            max_tokens=1500,  # 从2000降到1500
+            max_tokens=MAX_OUTPUT_TOKENS,
             temperature=0.5,
             stream=True
         )