167 lines
6.2 KiB
Python
167 lines
6.2 KiB
Python
import sys
|
|
import os
|
|
import asyncio
|
|
import nest_asyncio
|
|
import uuid
|
|
from agents import Runner, set_tracing_disabled, SQLiteSession
|
|
from openai.types.responses import ResponseTextDeltaEvent
|
|
|
|
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
|
|
|
import streamlit as st
|
|
|
|
from m12_agents_sdk_swarm.s01_tools import context_variables
|
|
from m12_agents_sdk_swarm.s02_agent import triage_agent, amap_server
|
|
|
|
nest_asyncio.apply()
|
|
set_tracing_disabled(True)
|
|
|
|
# 生成/获取动态 Session ID
|
|
if "session_id" not in st.session_state:
|
|
st.session_state.session_id = f"session_{uuid.uuid4().hex[:8]}"
|
|
|
|
|
|
async def init_mcp():
|
|
try:
|
|
await amap_server[0].connect()
|
|
return "✅ 高德地图（按需连接）"
|
|
except Exception as e:
|
|
return f"⚠️ MCP 连接失败"
|
|
|
|
|
|
mcp_status = asyncio.run(init_mcp())
|
|
|
|
st.set_page_config(page_title="智能客服驾驶舱", layout="wide")
|
|
st.title("✈️ 智能航空客服 Multi-Agent 系统")
|
|
st.caption(f"🚀 实战三：基于 Agents SDK 的多智能体协作 (ID: {st.session_state.session_id})") # 小字体显示文本
|
|
|
|
# 初始化持久化 Session
|
|
if "session" not in st.session_state:
|
|
st.session_state.session = SQLiteSession( # 持久化内存，将AI对话历史搬到数据库中
|
|
session_id=st.session_state.session_id,
|
|
db_path="./m13_streamlit/conversations.db"
|
|
)
|
|
|
|
if "display_messages" not in st.session_state: # 存储对话记录
|
|
st.session_state.display_messages = []
|
|
|
|
if "tool_logs_history" not in st.session_state: # 存储工具调用记录
|
|
st.session_state.tool_logs_history = []
|
|
|
|
|
|
# 封装侧边栏渲染函数，以便在初次加载和转接发生时，都能向同一个占位符刷新内容
|
|
def render_agent_status(placeholder, agent):
|
|
name = agent.name
|
|
if name == "TriageAgent":
|
|
placeholder.info("当前坐席: 前台 TriageAgent")
|
|
elif name == "RefundAgent":
|
|
placeholder.success("当前坐席: 退票专员 RefundAgent 🚨")
|
|
elif name == "ChangeAgent":
|
|
placeholder.warning("当前坐席: 改签专员 ChangeAgent 🔄")
|
|
|
|
|
|
with st.sidebar:
|
|
st.header("🖥️ 驾驶舱监控")
|
|
|
|
# 确保有大脑（session_state）记住当前 Agent
|
|
if "current_agent" not in st.session_state:
|
|
st.session_state.current_agent = triage_agent
|
|
|
|
# 当代码执行到对应位置时，自动更新上方侧边栏内容
|
|
agent_status_placeholder = st.empty()
|
|
|
|
# 初始渲染：基于当前state里的Agent
|
|
render_agent_status(agent_status_placeholder, st.session_state.current_agent)
|
|
|
|
st.subheader("👤 用户画像")
|
|
st.json(context_variables)
|
|
|
|
st.subheader("🔌 MCP 状态")
|
|
st.caption(mcp_status)
|
|
|
|
st.subheader("📊 会话统计")
|
|
st.metric("消息数", len(st.session_state.display_messages))
|
|
|
|
if st.button("🗑️ 清空对话"):
|
|
asyncio.run(st.session_state.session.clear_session()) # 调用SQLiteSession对象的clear_session的方法，删除所有session_id等于当前ID的记录
|
|
st.session_state.session_id = f"session_{uuid.uuid4().hex[:8]}" # 生成新的session_id
|
|
st.session_state.display_messages = [] # 重置UI缓存
|
|
st.session_state.tool_logs_history = []
|
|
st.session_state.current_agent = triage_agent
|
|
st.rerun()
|
|
|
|
# 渲染历史消息
|
|
for i, msg in enumerate(st.session_state.display_messages):
|
|
avatar = "👤" if msg["role"] == "user" else "🤖"
|
|
with st.chat_message(msg["role"], avatar=avatar):
|
|
st.write(msg["content"])
|
|
if i < len(st.session_state.tool_logs_history) and st.session_state.tool_logs_history[i]: # 确保列表索引存在且实际工具调用有记录
|
|
with st.expander("🔧 查看工具调用", expanded=False): # 默认折叠
|
|
for log in st.session_state.tool_logs_history[i]:
|
|
st.caption(log)
|
|
|
|
prompt = st.chat_input("请输入您的问题")
|
|
|
|
if prompt:
|
|
st.session_state.display_messages.append({"role": "user", "content": prompt})
|
|
st.session_state.tool_logs_history.append([]) # 提前占座，保证索引顺序
|
|
|
|
with st.chat_message("user", avatar="👤"): # 直接写入气泡
|
|
st.write(prompt)
|
|
|
|
with st.chat_message("assistant", avatar="🤖"):
|
|
message_placeholder = st.empty() # 先占位，AI思考后再填入
|
|
|
|
with st.status("Agent 正在思考...", expanded=True) as status:
|
|
async def process_stream():
|
|
stream = Runner.run_streamed(
|
|
st.session_state.current_agent,
|
|
input=prompt, # 问题
|
|
context=context_variables, # 全局上下文
|
|
session=st.session_state.session # 读:从数据库提取聊天记录/写:回复结束，自动将新一轮对话存入数据库
|
|
)
|
|
|
|
# 收集局部变量，再一起叠加到全局变量中
|
|
reply = "" # 回复内容累加器
|
|
tool_logs = [] # 工具日志收集
|
|
|
|
current_agent_name = st.session_state.current_agent.name
|
|
|
|
async for event in stream.stream_events(): # 流式事件
|
|
if event.type == "raw_response_event":
|
|
if isinstance(event.data, ResponseTextDeltaEvent):
|
|
delta = event.data.delta or "" # 提取新字符
|
|
reply += delta
|
|
message_placeholder.write(reply) # 实时刷新UI
|
|
|
|
# 当agent转换时
|
|
elif event.type == "agent_updated_stream_event":
|
|
new_agent = event.new_agent
|
|
if current_agent_name != new_agent.name:
|
|
log_msg = f"🔀 转接: {current_agent_name} → {new_agent.name}"
|
|
status.write(log_msg) # 写入折叠栏
|
|
tool_logs.append(log_msg)
|
|
|
|
# 1. 更新“大脑”（状态持久化）
|
|
st.session_state.current_agent = new_agent
|
|
# 2. 实时更新“脸面”（让侧边栏占位符立刻变色/变字）
|
|
render_agent_status(agent_status_placeholder, new_agent)
|
|
|
|
current_agent_name = new_agent.name
|
|
|
|
elif event.type == "run_item_stream_event":
|
|
if event.name == "tool_called":
|
|
tool_name = event.item.raw_item.name # 提取被调用工具的名字
|
|
log_msg = f"🔧 调用: {tool_name}"
|
|
status.write(log_msg) # 写入折叠栏
|
|
tool_logs.append(log_msg)
|
|
|
|
return reply, tool_logs # 返回完整回复的字符串与收集到的动作日志
|
|
|
|
|
|
reply, tool_logs = asyncio.run(process_stream())
|
|
status.update(label="✅ 处理完成", state="complete", expanded=False) # 标识处理成功
|
|
|
|
# 将(本轮)对话记录/工具调用，分别存入(总)对话记录/工具调用
|
|
st.session_state.display_messages.append({"role": "assistant", "content": reply})
|
|
st.session_state.tool_logs_history.append(tool_logs) |