FiDA_Python/src/server/agent/graph.py

import random
from typing import Literal
from langchain_core.messages import AIMessage
from langchain_core.runnables import RunnableConfig
from langchain_qwq import ChatQwen
from langgraph.checkpoint.serde.jsonplus import JsonPlusSerializer
from langgraph.graph import StateGraph, END, START
from pydantic import BaseModel
from pymongo import MongoClient

from src.core.config import MONGO_URI, settings
from src.server.agent.state import AgentState
from src.server.agent.agents import designer_node, researcher_node, visualizer_node, suggester_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):
    # 将 FINISH 替换或增加 Suggester
    next: Literal["Designer", "Researcher", "Visualizer", "Suggester", "FINISH"]


llm_supervisor = ChatQwen(
    model="qwen3.5-flash",
    max_tokens=3_000,
    timeout=None,
    max_retries=2,
    api_key=settings.QWEN_API_KEY)


def supervisor_node(state: AgentState, config: RunnableConfig):
    configurable = config["configurable"]
    use_report = configurable.get("use_report", False)
    suggest_frequency = configurable.get("require_suggestion", 0.6)  # 0.0~1.0

    messages = state["messages"]
    if not messages:
        return {"next": "Suggester"}

    # ── system prompt 保持不变 ──
    system_prompt = f"""你是家具设计主管，负责分配任务。
当前设定：
- 是否需要市场研究报告：{'是' if use_report else '否'}

严格遵守以下规则：
- 如果 **不需要** 市场研究报告（use_report = False），**绝对不能** 选择 Researcher
- 只有在 **明确需要** 市场报告、竞争分析、材质趋势、价格区间等外部资讯时，才选择 Researcher，且 **必须** use_report = True
- 常见分配：
  - 纯设计、风格、尺寸、材质建议 → Designer
  - 需要生成图片、渲染 → Visualizer
  - 需要产生建议按钮 → Suggester
  - 需要市场报告 → Researcher（但只有 use_report=True 时才允许）
  - 对话已完整、无需继续 → FINISH

用户最后说了什么？请根据实际需求决定下一步。
"""

    chain = llm_supervisor.with_structured_output(RouteResponse)
    decision = chain.invoke([{"role": "system", "content": system_prompt}] + messages)
    next_node = decision.next  # 防空默认 FINISH

    # 安全阀：禁止非法选择 Researcher
    if next_node == "Researcher" and not use_report:
        print("警告：LLM 违规选择了 Researcher，已强制改为 Suggester 或 FINISH")
        next_node = "Suggester" if state.get("require_suggestion", False) else "FINISH"

    # 核心改动：只有 LLM 决定 FINISH 时，才掷骰子看是否插入 Suggester
    if next_node == "FINISH":
        # 满足概率条件 → 插入 Suggester
        if suggest_frequency > 0 and random.random() < suggest_frequency:
            next_node = "Suggester"

    return {"next": next_node}


# --- 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_node("Suggester", suggester_node)
workflow.add_edge(START, "Supervisor")

# 修改条件边映射
workflow.add_conditional_edges(
    "Supervisor",
    lambda state: state["next"],
    {
        "Designer": "Designer",
        "Researcher": "Researcher",
        "Visualizer": "Visualizer",
        "Suggester": "Suggester",  # 原本的 FINISH 现在指向 Suggester
        "FINISH": END  # 直接结束，不给建议
    }
)

# 专家执行完依然回到 Supervisor
workflow.add_edge("Designer", "Supervisor")
workflow.add_edge("Researcher", "Supervisor")
workflow.add_edge("Visualizer", "Supervisor")
# 重点：Suggester 可以是整个流程的终点
workflow.add_edge("Suggester", END)

client = MongoClient(MONGO_URI)
checkpointer = MongoDBSaver(
    client=client["furniture_agent_db"],
    db_name="langgraph",
    collection_name="checkpoints",
    serde=JsonPlusSerializer(pickle_fallback=True),  # ← 關鍵這一行
)
app = workflow.compile(checkpointer=checkpointer)