root
b5c15ef445
All checks were successful
构建并部署 AI Agent 服务 / deploy (push) Successful in 12m9s
## 核心改动 ### 1. 单图方案重构 - 删除了多图（self.graphs），改为单图（self.graph） - 新增 MainGraphState.current_model 字段用于运行时注入模型 - llm_call 节点改为动态选择模型（create_dynamic_llm_call_node） ### 2. chat_services 优化 - 添加 _cached_services 缓存，避免重复初始化 - 新增 get_cached_chat_services() 函数，用于单图注入 - 新增 _check_http_service_available() 统一HTTP探测逻辑 - 减少重复代码，LocalVLLMChatProvider和LocalSmallModelProvider共用探测方法 ### 3. AIAgentService 重构 - initialize() 只构建一次图，传入 chat_services 字典 - 新增 _resolve_model() 模型回退逻辑 - 新增 _build_invocation() 统一构建调用参数 - process_message() 和 process_message_stream() 改为注入 current_model - 流式处理代码拆分，增加可读性 ### 4. 新增和删除文件 - 新增：backend/app/main_graph/main_graph_builder.py（图构建） - 新增：backend/app/main_graph/subgraph_wrapper.py（子图封装） - 新增：tools/test/test_tavily_search.py（测试） - 删除：backend/app/main_graph/graph.py（旧图） - 删除：backend/app/main_graph/utils/main_graph_builder.py（旧构建器） - 删除：backend/app/main_graph/utils/__init__.py ### 5. 其他更新 - README.md：新增模型服务使用情况详解章节 - backend/app/model_services/__init__.py：新增 get_cached_chat_services 导出 ## 方案优势 - 内存优化：N张图 → 1张图 - 灵活性：运行时动态选择模型，支持同会话不同模型 - 性能：模型服务缓存，初始化仅一次 - 可维护性：减少重复代码，统一HTTP探测逻辑
207 lines
8.6 KiB
Python
207 lines
8.6 KiB
Python
"""
|
||
LLM 调用节点模块
|
||
负责调用大语言模型并处理响应
|
||
"""
|
||
|
||
import time
|
||
from typing import Any, Dict
|
||
from langchain_core.language_models import BaseChatModel
|
||
from langchain_core.messages import AIMessage
|
||
|
||
# 本地模块
|
||
from app.main_graph.state import MainGraphState
|
||
from app.agent.prompts import create_system_prompt
|
||
from app.utils.logging import log_state_change
|
||
from app.logger import debug, info, error
|
||
|
||
|
||
def create_dynamic_llm_call_node(chat_services: Dict[str, BaseChatModel], tools: list):
|
||
"""
|
||
工厂函数：创建动态 LLM 调用节点（根据 state.current_model 选择模型）
|
||
|
||
Args:
|
||
chat_services: 模型名称 -> ChatModel 实例 的字典
|
||
tools: 工具列表
|
||
|
||
Returns:
|
||
异步节点函数
|
||
"""
|
||
# 预构建所有模型的 tools 绑定（避免每次调用都 bind）
|
||
bound_models: Dict[str, Any] = {}
|
||
for name, llm in chat_services.items():
|
||
if tools:
|
||
bound_models[name] = llm.bind_tools(tools)
|
||
else:
|
||
bound_models[name] = llm
|
||
|
||
# 预构建 prompt
|
||
prompt = create_system_prompt(tools)
|
||
|
||
from langchain_core.runnables.config import RunnableConfig
|
||
|
||
async def call_llm(state: MainGraphState, config: RunnableConfig) -> Dict[str, Any]:
|
||
"""
|
||
LLM 调用节点（动态选择模型）
|
||
|
||
Args:
|
||
state: 当前对话状态
|
||
config: LangChain/LangGraph 自动注入的配置，包含 callbacks 等信息
|
||
|
||
Returns:
|
||
更新后的状态字典
|
||
"""
|
||
log_state_change("llm_call", state, "进入")
|
||
|
||
memory_context = getattr(state, "memory_context", "暂无用户信息")
|
||
start_time = time.time()
|
||
|
||
# 关键修复：如果 state.final_result 已经存在（比如子图执行完），直接返回
|
||
if state.final_result:
|
||
info(f"[llm_call] 检测到已有最终结果，直接返回: {state.final_result[:100]}...")
|
||
elapsed_time = time.time() - start_time
|
||
return {
|
||
"final_result": state.final_result,
|
||
"success": True,
|
||
"current_phase": "done",
|
||
"llm_calls": getattr(state, 'llm_calls', 0) + 1,
|
||
"last_elapsed_time": elapsed_time,
|
||
"turns_since_last_summary": getattr(state, 'turns_since_last_summary', 0) + 1,
|
||
}
|
||
|
||
# 动态选择模型
|
||
model_name = getattr(state, "current_model", "")
|
||
if not model_name or model_name not in bound_models:
|
||
# 回退到第一个可用模型
|
||
fallback_name = next(iter(bound_models.keys()))
|
||
info(f"[llm_call] 模型 '{model_name}' 不可用，回退到 '{fallback_name}'")
|
||
model_name = fallback_name
|
||
|
||
llm_with_tools = bound_models[model_name]
|
||
info(f"[llm_call] 使用模型: {model_name}")
|
||
|
||
try:
|
||
# 添加上下文到消息
|
||
messages_with_context = list(state.messages)
|
||
if state.rag_context:
|
||
from langchain_core.messages import SystemMessage
|
||
rag_system_msg = SystemMessage(content=f"以下是检索到的相关信息：\n{state.rag_context}")
|
||
inserted = False
|
||
for i, msg in enumerate(messages_with_context):
|
||
if msg.type == "human":
|
||
messages_with_context.insert(i, rag_system_msg)
|
||
inserted = True
|
||
break
|
||
if not inserted:
|
||
messages_with_context.insert(0, rag_system_msg)
|
||
|
||
# 恢复为：手动进行 astream，并将所有的 chunk 拼接成最终的 response 返回。
|
||
# LangGraph 会自动监听这期间产生的所有 token。
|
||
chain = prompt | llm_with_tools
|
||
chunks = []
|
||
async for chunk in chain.astream(
|
||
{
|
||
"messages": messages_with_context,
|
||
"memory_context": memory_context
|
||
},
|
||
config=config
|
||
):
|
||
chunks.append(chunk)
|
||
|
||
# 将所有 chunk 合并成最终的 AIMessage
|
||
if chunks:
|
||
response = chunks[0]
|
||
for chunk in chunks[1:]:
|
||
response = response + chunk
|
||
else:
|
||
response = AIMessage(content="")
|
||
|
||
elapsed_time = time.time() - start_time
|
||
|
||
# 提取 token 用量（兼容不同 LLM 提供商的元数据格式）
|
||
token_usage = {}
|
||
input_tokens = 0
|
||
output_tokens = 0
|
||
|
||
# 尝试从 response_metadata 中提取
|
||
if hasattr(response, 'response_metadata') and response.response_metadata:
|
||
meta = response.response_metadata
|
||
if 'token_usage' in meta:
|
||
token_usage = meta['token_usage']
|
||
elif 'usage' in meta:
|
||
token_usage = meta['usage']
|
||
|
||
# 尝试从 additional_kwargs 中提取
|
||
if not token_usage and hasattr(response, 'additional_kwargs'):
|
||
add_kwargs = response.additional_kwargs
|
||
if 'llm_output' in add_kwargs and 'token_usage' in add_kwargs['llm_output']:
|
||
token_usage = add_kwargs['llm_output']['token_usage']
|
||
|
||
# 提取具体的 token 数值
|
||
if token_usage:
|
||
input_tokens = token_usage.get('prompt_tokens', token_usage.get('input_tokens', 0))
|
||
output_tokens = token_usage.get('completion_tokens', token_usage.get('output_tokens', 0))
|
||
|
||
# 打印 LLM 的完整输出
|
||
debug("\n" + "="*80)
|
||
debug(f"📥 [LLM输出] 模型: {model_name} 返回的完整响应:")
|
||
debug(f" 消息类型: {response.type.upper()}")
|
||
debug(f" 内容长度: {len(str(response.content))} 字符")
|
||
debug("-"*80)
|
||
debug(f"{response.content}")
|
||
|
||
# 打印响应统计信息
|
||
info(f"⏱️ [LLM统计] 调用耗时: {elapsed_time:.2f}秒")
|
||
info(f"📊 [LLM统计] Token用量: 输入={input_tokens}, 输出={output_tokens}, 总计={input_tokens + output_tokens}")
|
||
if token_usage:
|
||
debug(f"📋 [LLM统计] 详细用量: {token_usage}")
|
||
debug("="*80 + "\n")
|
||
|
||
# 检查是否有工具调用
|
||
has_tool_calls = hasattr(response, 'tool_calls') and len(response.tool_calls) > 0
|
||
|
||
result = {
|
||
"messages": [response],
|
||
"llm_calls": getattr(state, 'llm_calls', 0) + 1,
|
||
"last_token_usage": token_usage,
|
||
"last_elapsed_time": elapsed_time,
|
||
"turns_since_last_summary": getattr(state, 'turns_since_last_summary', 0) + 1,
|
||
"final_result": response.content,
|
||
"success": True,
|
||
"current_phase": "done",
|
||
"has_tool_calls": has_tool_calls,
|
||
"current_model": model_name # 记录实际使用的模型
|
||
}
|
||
|
||
log_state_change("llm_call", state, "离开")
|
||
return result
|
||
|
||
except Exception as e:
|
||
elapsed_time = time.time() - start_time
|
||
error(f"\n❌ [LLM错误] 模型 {model_name} 调用失败 (耗时: {elapsed_time:.2f}秒)")
|
||
error(f" 错误类型: {type(e).__name__}")
|
||
error(f" 错误信息: {str(e)}")
|
||
import traceback
|
||
error(f"📋 堆栈: {traceback.format_exc()}")
|
||
debug("="*80 + "\n")
|
||
|
||
# 返回一个友好的错误消息
|
||
error_response = AIMessage(
|
||
content="抱歉，模型暂时无法响应，可能是网络超时或服务繁忙，请稍后再试。"
|
||
)
|
||
error_result = {
|
||
"messages": [error_response],
|
||
"llm_calls": getattr(state, 'llm_calls', 0),
|
||
"last_token_usage": {},
|
||
"last_elapsed_time": elapsed_time,
|
||
"turns_since_last_summary": getattr(state, 'turns_since_last_summary', 0) + 1,
|
||
"final_result": "抱歉，模型暂时无法响应，可能是网络超时或服务繁忙，请稍后再试。",
|
||
"success": False,
|
||
"current_phase": "done",
|
||
"current_model": model_name
|
||
}
|
||
|
||
log_state_change("llm_call", state, "离开（异常）")
|
||
return error_result
|
||
|
||
return call_llm
|