| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303 |
- # app/services/agent/engine.py
- import json
- import uuid
- import asyncio
- from typing import AsyncGenerator, Dict, Any, List, Optional
- from sqlalchemy.ext.asyncio import AsyncSession
- from redis.asyncio import Redis
- from app.core.logger import logger
- from app.models.user import VasUser
- from app.schemas.agent import (
- AgentChatRequest,
- AgentChatResponse,
- AgentToolCallStep,
- )
- from app.services.llm_service import LlmService
- from app.services.agent.memory import AgentMemory
- from app.services.agent.tools.registry import tool_registry
- # 自动加载内置业务工具
- import app.services.agent.tools.builtin # noqa: F401
- SYSTEM_PROMPT = """你是一个专业的智能业务助理,具备自主调用系统后端工具(Tools)的能力。
- 你可以根据用户的自然语言需求,自主规划并按顺序调用相关工具来完成查询、数据检索或操作任务。
- 【行为规范与准则】:
- 1. 遇到需要查询、修改、统计具体业务数据时,必须调用对应的工具获取真实结果,禁止凭空捏造数据或订单状态。
- 2. 若用户需要执行操作(如取消订单),请先确认订单状态及所需参数,必要时先查询详情再执行操作。
- 3. 如果工具返回错误或未查到结果,请友好告知用户原因或提出可能的解决建议。
- 4. 工具调用完成后,请用清晰、结构化、自然的语言向用户总结执行结果。
- 5. 回答始终保持礼貌、高效与专业。"""
- class AgentEngine:
- """Agent 核心编排引擎,支持 ReAct 循环与 SSE 流式输出"""
- @classmethod
- def _format_sse(cls, event: str, data: Dict[str, Any]) -> str:
- return f"event: {event}\ndata: {json.dumps(data, ensure_ascii=False)}\n\n"
- @classmethod
- async def run_sync(
- cls,
- request: AgentChatRequest,
- current_user: VasUser,
- db: AsyncSession,
- redis_client: Optional[Redis] = None,
- ) -> AgentChatResponse:
- """同步执行 Agent 对话并返回完整结果"""
- session_id = request.session_id or f"sess_{uuid.uuid4().hex[:12]}"
- history = await AgentMemory.get_history(session_id, max_messages=10)
- tools_schema = tool_registry.get_openai_tools_schema()
- context = {
- "db": db,
- "current_user": current_user,
- "redis_client": redis_client,
- }
- # 组装本轮执行的消息列表
- messages: List[Dict[str, Any]] = [
- {"role": "system", "content": SYSTEM_PROMPT},
- ]
- # 追加历史对话
- messages.extend(history)
- # 追加当前用户输入
- user_msg = {"role": "user", "content": request.message}
- messages.append(user_msg)
- # 用于持久化到 Redis 的增量消息
- new_history_messages: List[Dict[str, Any]] = [user_msg]
- steps_record: List[AgentToolCallStep] = []
- step = 0
- final_answer = ""
- while step < request.max_steps:
- step += 1
- logger.info(f"[AgentEngine] Step {step}/{request.max_steps} calling LLM (session={session_id})...")
- llm_reply = await LlmService.chat_completion_with_tools_async(
- messages=messages,
- tools=tools_schema if tools_schema else None,
- model=request.model,
- )
- tool_calls = llm_reply.get("tool_calls")
- content = llm_reply.get("content") or ""
- if not tool_calls:
- # LLM 没有触发工具调用,直接作为最终答案返回
- final_answer = content
- assistant_msg = {"role": "assistant", "content": final_answer}
- new_history_messages.append(assistant_msg)
- break
- # 处理工具调用
- assistant_call_msg = {
- "role": "assistant",
- "content": content,
- "tool_calls": tool_calls,
- }
- messages.append(assistant_call_msg)
- new_history_messages.append(assistant_call_msg)
- for call in tool_calls:
- fn_info = call.get("function", {})
- tool_name = fn_info.get("name", "")
- raw_args = fn_info.get("arguments", "{}")
- call_id = call.get("id", f"call_{uuid.uuid4().hex[:8]}")
- try:
- args = json.loads(raw_args) if isinstance(raw_args, str) else raw_args
- except Exception:
- args = {}
- logger.info(f"[AgentEngine] Invoking tool: {tool_name} with args: {args}")
- tool_result = await tool_registry.execute(tool_name, args, context=context)
- is_error = isinstance(tool_result, dict) and "error" in tool_result
- steps_record.append(
- AgentToolCallStep(
- tool=tool_name,
- arguments=args,
- result=tool_result,
- is_error=is_error,
- )
- )
- tool_reply_msg = {
- "role": "tool",
- "tool_call_id": call_id,
- "name": tool_name,
- "content": json.dumps(tool_result, ensure_ascii=False),
- }
- messages.append(tool_reply_msg)
- new_history_messages.append(tool_reply_msg)
- if not final_answer and step >= request.max_steps:
- # 达到最大步数未停止,强制做一次总结
- summary_reply = await LlmService.chat_completion_with_tools_async(
- messages=messages + [{"role": "user", "content": "请根据以上所有工具执行结果,做最终汇总回答。"}],
- tools=None,
- model=request.model,
- )
- final_answer = summary_reply.get("content", "已完成操作处理。")
- new_history_messages.append({"role": "assistant", "content": final_answer})
- # 持久化更新历史记录到 Redis(只保存完整问答对,避免孤立 tool 消息和上下文暴涨)
- if final_answer:
- await AgentMemory.append_messages(session_id, [
- {"role": "user", "content": request.message},
- {"role": "assistant", "content": final_answer},
- ])
- return AgentChatResponse(
- session_id=session_id,
- answer=final_answer,
- steps=steps_record,
- steps_count=len(steps_record),
- )
- @classmethod
- async def run_stream(
- cls,
- request: AgentChatRequest,
- current_user: VasUser,
- db: AsyncSession,
- redis_client: Optional[Redis] = None,
- ) -> AsyncGenerator[str, None]:
- """以 SSE 格式流式生成 Agent 运行状态、工具调用过程与最终回答"""
- session_id = request.session_id or f"sess_{uuid.uuid4().hex[:12]}"
- try:
- yield cls._format_sse("think", {"status": "started", "session_id": session_id})
- history = await AgentMemory.get_history(session_id, max_messages=10)
- tools_schema = tool_registry.get_openai_tools_schema()
- context = {
- "db": db,
- "current_user": current_user,
- "redis_client": redis_client,
- }
- messages: List[Dict[str, Any]] = [
- {"role": "system", "content": SYSTEM_PROMPT},
- ]
- messages.extend(history)
- user_msg = {"role": "user", "content": request.message}
- messages.append(user_msg)
- new_history_messages: List[Dict[str, Any]] = [user_msg]
- steps_record: List[AgentToolCallStep] = []
- step = 0
- final_answer = ""
- while step < request.max_steps:
- step += 1
- yield cls._format_sse("think", {"status": "planning", "step": step})
- llm_reply = await LlmService.chat_completion_with_tools_async(
- messages=messages,
- tools=tools_schema if tools_schema else None,
- model=request.model,
- )
- tool_calls = llm_reply.get("tool_calls")
- content = llm_reply.get("content") or ""
- reasoning = llm_reply.get("reasoning_content") or ""
- if reasoning:
- yield cls._format_sse("think", {"status": "reasoning", "reasoning": reasoning, "step": step})
- if not tool_calls:
- final_answer = content
- # 推送最终回复文本
- yield cls._format_sse("message", {"delta": final_answer, "full": final_answer})
- new_history_messages.append({"role": "assistant", "content": final_answer})
- break
- # 记录 assistant 的 tool_calls
- assistant_call_msg = {
- "role": "assistant",
- "content": content,
- "tool_calls": tool_calls,
- }
- messages.append(assistant_call_msg)
- new_history_messages.append(assistant_call_msg)
- for call in tool_calls:
- fn_info = call.get("function", {})
- tool_name = fn_info.get("name", "")
- raw_args = fn_info.get("arguments", "{}")
- call_id = call.get("id", f"call_{uuid.uuid4().hex[:8]}")
- try:
- args = json.loads(raw_args) if isinstance(raw_args, str) else raw_args
- except Exception:
- args = {}
- # 推送 tool_start 事件
- yield cls._format_sse("tool_start", {
- "tool": tool_name,
- "arguments": args,
- "call_id": call_id,
- })
- tool_result = await tool_registry.execute(tool_name, args, context=context)
- is_error = isinstance(tool_result, dict) and "error" in tool_result
- step_obj = AgentToolCallStep(
- tool=tool_name,
- arguments=args,
- result=tool_result,
- is_error=is_error,
- )
- steps_record.append(step_obj)
- # 推送 tool_end 事件
- yield cls._format_sse("tool_end", {
- "tool": tool_name,
- "result": tool_result,
- "is_error": is_error,
- "call_id": call_id,
- })
- tool_reply_msg = {
- "role": "tool",
- "tool_call_id": call_id,
- "name": tool_name,
- "content": json.dumps(tool_result, ensure_ascii=False),
- }
- messages.append(tool_reply_msg)
- new_history_messages.append(tool_reply_msg)
- if not final_answer and step >= request.max_steps:
- summary_reply = await LlmService.chat_completion_with_tools_async(
- messages=messages + [{"role": "user", "content": "请根据以上所有工具执行结果,做最终汇总回答。"}],
- tools=None,
- model=request.model,
- )
- final_answer = summary_reply.get("content", "已完成操作处理。")
- yield cls._format_sse("message", {"delta": final_answer, "full": final_answer})
- new_history_messages.append({"role": "assistant", "content": final_answer})
- # 更新记忆(只保存完整问答对,避免孤立 tool 消息和跨轮上下文结构错乱)
- if final_answer:
- await AgentMemory.append_messages(session_id, [
- {"role": "user", "content": request.message},
- {"role": "assistant", "content": final_answer},
- ])
- # 发送 finish 完成事件
- yield cls._format_sse("finish", {
- "session_id": session_id,
- "answer": final_answer,
- "steps_count": len(steps_record),
- })
- except Exception as e:
- logger.error(f"[AgentEngine.stream] error: {e}", exc_info=True)
- yield cls._format_sse("error", {"error": str(e)})
|