feat: v3.0 Function Calling模式 - LLM自主调用工具

services/llm_service.py

@@ -382,5 +382,200 @@ class LLMService:
             yield {"type": "content", "text": buffer}
 
 
+    async def chat_with_tools(
+        self,
+        messages: List[Dict],
+        provider_config: dict,
+        agent_config: dict,
+        tools: List[Dict] = None,
+        enable_thinking: bool = True,
+        images: List[Dict] = None
+    ) -> Tuple[str, Optional[str], Optional[List[Dict]]]:
+        """
+        支持Function Calling的对话
+        
+        Args:
+            messages: 对话历史
+            provider_config: LLM Provider配置
+            agent_config: Agent配置
+            tools: 工具定义列表（OpenAI Function Calling格式）
+            enable_thinking: 是否启用思考
+            images: 图片数据列表
+        
+        Returns:
+            Tuple[str, Optional[str], Optional[List[Dict]]]: (回复内容, 思考过程, 工具调用记录)
+        """
+        api_base = provider_config.get('api_base')
+        api_key = provider_config.get('api_key')
+        model = agent_config.get('model_override') or provider_config.get('default_model', 'auto')
+        supports_function_calling = provider_config.get('supports_function_calling', False)
+        
+        max_tokens = provider_config.get('max_tokens', 4096)
+        temperature = agent_config.get('temperature_override') or provider_config.get('temperature', 0.7)
+        
+        # 如果不支持Function Calling，直接调用普通chat
+        if not supports_function_calling or not tools:
+            response, thinking = await self.chat(messages, provider_config, agent_config, enable_thinking, images)
+            return response, thinking, None
+        
+        # 构建消息
+        final_messages = messages.copy()
+        system_prompt = agent_config.get('system_prompt', '你是一个有用的AI助手。')
+        if final_messages and final_messages[0]['role'] != 'system':
+            final_messages.insert(0, {"role": "system", "content": system_prompt})
+        
+        # 处理图片（多模态）
+        if images and len(images) > 0:
+            for i in range(len(final_messages) - 1, -1, -1):
+                if final_messages[i]['role'] == 'user':
+                    original_text = final_messages[i]['content']
+                    multimodal_content = [{"type": "text", "text": original_text if original_text else "请描述这张图片"}]
+                    for img in images:
+                        multimodal_content.append({
+                            "type": "image_url",
+                            "image_url": {"url": img['data']}
+                        })
+                    final_messages[i]['content'] = multimodal_content
+                    break
+        
+        # 第一次调用：让LLM决定是否调用工具
+        url = f"{api_base.rstrip('/')}/chat/completions"
+        headers = {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json"
+        }
+        payload = {
+            "model": model,
+            "messages": final_messages,
+            "temperature": temperature,
+            "max_tokens": max_tokens,
+            "tools": tools  # 传入工具定义
+        }
+        
+        logger.info(f"Function Calling调用: url={url}, model={model}, tools={len(tools)}")
+        
+        tool_calls_record = []  # 记录工具调用
+        
+        try:
+            async with httpx.AsyncClient(timeout=60.0) as client:
+                response = await client.post(url, headers=headers, json=payload)
+                
+                if response.status_code != 200:
+                    logger.error(f"API返回错误: status={response.status_code}, body={response.text[:500]}")
+                    response.raise_for_status()
+                
+                data = response.json()
+                
+                if 'choices' not in data or len(data['choices']) == 0:
+                    raise ValueError("API响应格式错误：缺少choices")
+                
+                message = data['choices'][0]['message']
+                
+                # 检查是否有工具调用
+                if 'tool_calls' in message and message['tool_calls']:
+                    logger.info(f"LLM请求调用工具: {len(message['tool_calls'])} 个")
+                    
+                    # 将LLM的工具调用消息添加到历史
+                    final_messages.append({
+                        "role": "assistant",
+                        "content": None,
+                        "tool_calls": message['tool_calls']
+                    })
+                    
+                    # 记录工具调用
+                    for tc in message['tool_calls']:
+                        tool_calls_record.append({
+                            "id": tc['id'],
+                            "name": tc['function']['name'],
+                            "arguments": json.loads(tc['function']['arguments'])
+                        })
+                    
+                    # 返回工具调用记录，由调用方执行工具
+                    return None, None, tool_calls_record
+                
+                # 没有工具调用，直接返回内容
+                content = message.get('content', '')
+                
+                # 处理思考内容（如果有）
+                thinking_content = None
+                # 这里可以添加思考内容提取逻辑
+                
+                return content, thinking_content, None
+                
+        except httpx.HTTPStatusError as e:
+            logger.error(f"HTTP错误: {e.response.status_code}, {e.response.text}")
+            raise
+        except Exception as e:
+            logger.error(f"Function Calling调用异常: {type(e).__name__}: {e}")
+            raise
+    
+    async def chat_with_tool_results(
+        self,
+        messages: List[Dict],
+        provider_config: dict,
+        agent_config: dict,
+        tool_results: List[Dict],
+        enable_thinking: bool = True
+    ) -> Tuple[str, Optional[str]]:
+        """
+        第二阶段调用：将工具执行结果返回给LLM
+        
+        Args:
+            messages: 对话历史（包含工具调用和结果）
+            provider_config: LLM Provider配置
+            agent_config: Agent配置
+            tool_results: 工具执行结果 [{"tool_call_id": "xxx", "content": "..."}]
+        
+        Returns:
+            Tuple[str, Optional[str]]: (回复内容, 思考过程)
+        """
+        api_base = provider_config.get('api_base')
+        api_key = provider_config.get('api_key')
+        model = agent_config.get('model_override') or provider_config.get('default_model', 'auto')
+        max_tokens = provider_config.get('max_tokens', 4096)
+        temperature = agent_config.get('temperature_override') or provider_config.get('temperature', 0.7)
+        
+        # 将工具结果添加到消息历史
+        final_messages = messages.copy()
+        for result in tool_results:
+            final_messages.append({
+                "role": "tool",
+                "tool_call_id": result['tool_call_id'],
+                "content": result['content']
+            })
+        
+        # 调用LLM生成最终回复
+        url = f"{api_base.rstrip('/')}/chat/completions"
+        headers = {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json"
+        }
+        payload = {
+            "model": model,
+            "messages": final_messages,
+            "temperature": temperature,
+            "max_tokens": max_tokens
+        }
+        
+        logger.info(f"工具结果返回LLM: url={url}, model={model}")
+        
+        try:
+            async with httpx.AsyncClient(timeout=60.0) as client:
+                response = await client.post(url, headers=headers, json=payload)
+                
+                if response.status_code != 200:
+                    logger.error(f"API返回错误: status={response.status_code}")
+                    response.raise_for_status()
+                
+                data = response.json()
+                content = data['choices'][0]['message']['content']
+                
+                return content, None
+                
+        except Exception as e:
+            logger.error(f"工具结果调用异常: {e}")
+            raise
+
+
 # 全局实例
 llm_service = LLMService()