first commit

src/server/agent/graph.py

@@ -0,0 +1,98 @@
+import os
+from typing import Literal
+
+from google.oauth2 import service_account
+from langchain_core.messages import AIMessage
+from langchain_google_genai import ChatGoogleGenerativeAI
+from langgraph.graph import StateGraph, END, START
+from pydantic import BaseModel
+from pymongo import MongoClient
+
+from src.core.config import settings, MONGO_URI
+from src.server.agent.state import AgentState
+from src.server.agent.agents import designer_node, researcher_node, visualizer_node
+from langgraph.checkpoint.mongodb import MongoDBSaver
+
+
+# --- Supervisor (路由逻辑) ---
+# 定义路由的输出结构，强制 LLM 选择一个
+class RouteResponse(BaseModel):
+    next: Literal["Designer", "Researcher", "Visualizer", "FINISH"]
+
+
+creds = service_account.Credentials.from_service_account_file(
+    settings.GOOGLE_GENAI_USE_VERTEXAI,
+    scopes=["https://www.googleapis.com/auth/cloud-platform"],
+)
+
+llm_supervisor = ChatGoogleGenerativeAI(
+    model="gemini-2.0-flash", temperature=0, credentials=creds,
+    project="aida-461108", location='us-central1', vertexai=True, api_key=settings.GOOGLE_API_KEY
+)
+
+
+def supervisor_node(state: AgentState):
+    messages = state["messages"]
+    if not messages:
+        return {"next": "FINISH"}
+
+    last_message = messages[-1]
+
+    # --- 改进的拦截逻辑 ---
+    # 如果最后一条消息是 AI 产生的（且没有调用工具），说明专家已经回复完了用户
+    # 此时我们才拦截并结束，否则会导致专家没机会说话
+    if isinstance(last_message, AIMessage) and not last_message.tool_calls:
+        return {"next": "FINISH"}
+
+    # 如果最后一条是 HumanMessage，说明用户刚说完，Supervisor 必须派发任务
+    system_prompt = """
+    你是家具设计团队的主管(Supervisor)。
+    请根据用户的意图，选择最合适的专家：
+    - Designer: 设计建议、参数细化、闲聊、问候。
+    - Visualizer: 绘图、看草图。
+    - Researcher: 市场报告、趋势。
+
+    只需输出专家名称。
+    """
+
+    chain = llm_supervisor.with_structured_output(RouteResponse)
+    decision = chain.invoke([{"role": "system", "content": system_prompt}] + messages)
+
+    return {"next": decision.next}
+
+
+# --- 构建 Graph ---
+workflow = StateGraph(AgentState)
+
+workflow.add_node("Supervisor", supervisor_node)
+workflow.add_node("Designer", designer_node)
+workflow.add_node("Researcher", researcher_node)
+workflow.add_node("Visualizer", visualizer_node)
+
+workflow.add_edge(START, "Supervisor")
+
+# 这里的逻辑是关键：Supervisor 决定去向
+workflow.add_conditional_edges(
+    "Supervisor",
+    lambda state: state["next"],
+    {
+        "Designer": "Designer",
+        "Researcher": "Researcher",
+        "Visualizer": "Visualizer",
+        "FINISH": END
+    }
+)
+
+# 重点修改：专家执行完后，必须回到 Supervisor 进行状态检查
+# 如果 Supervisor 发现专家刚说完话，它会触发上面的逻辑返回 FINISH
+workflow.add_edge("Designer", "Supervisor")
+workflow.add_edge("Researcher", "Supervisor")
+workflow.add_edge("Visualizer", "Supervisor")
+
+client = MongoClient(MONGO_URI)
+checkpointer = MongoDBSaver(
+    client=client["furniture_agent_db"],
+    db_name="langgraph",
+    collection_name="checkpoints"
+)
+app = workflow.compile(checkpointer=checkpointer)