cfc/cfc-langgraph/app/api/meal.py

import uuid
import json
import logging
import re
from fastapi import APIRouter
from app.models.meal import (
    FoodRecognizeRequest,
    FoodRecognizeResponse,
    FoodsItem,
    MenuGenerateRequest,
    MenuGenerateResponse,
)
from app.config import settings
from app.prompt_service import get_prompt

logger = logging.getLogger(__name__)
router = APIRouter(prefix="/api/v1", tags=["meal"])

FOOD_RECOGNIZE_TEMPLATE = """请识别以下图片中的主要食材（3-8种），按置信度排序。
返回 JSON 数组格式，不要包含其他内容：
[{{"name": "食材名称", "confidence": 0.95, "category": "蔬菜/水果/肉禽/水产/蛋奶/谷物/调味/其他"}}]

图片链接: {image_url}
要求: 只返回 JSON 数组，不要有任何其他文字。"""

MENU_GENERATE_TEMPLATE = """请为{participant_count}人生成{date}的一日三餐菜单。

可用食材: {foods_text}
用餐人数: {participant_count}人
{constraints_text}

请返回 JSON 格式：
{{"meals": [{{"type": "breakfast", "name": "早餐", "dishes": [{{"name": "菜品名", "ingredients": [{{"name": "食材", "grams": 100}}], "cooking_method": "做法一句话", "nutrition": {{"calories": 200}}}}]}}]}}

要求:
1. 午餐和晚餐按上方约束给出的菜数配菜（D(N)=4+max(0,N-3) 道菜 + 1 道汤，汤恒 1 且不计入菜数、只加菜不加汤），早餐维持 1-2 道
2. 食材用量按{participant_count}人份计算
3. 营养均衡，考虑健康目标
4. 只用提供的食材
5. 每道菜用 type 字段标注 dish（菜）或 soup（汤），汤恰好 1 道且排在最后
6. 只返回 JSON，不要其他文字"""


@router.post("/food/recognize", response_model=FoodRecognizeResponse)
async def recognize_food(req: FoodRecognizeRequest):
    """食材识别: 上传图片 URL, 返回识别食材列表"""
    trace_id = str(uuid.uuid4())
    try:
        from langchain_openai import ChatOpenAI
        from langchain_core.messages import HumanMessage, SystemMessage

        llm = ChatOpenAI(
            model=settings.llm_model,
            api_key=settings.llm_api_key,
            base_url=settings.llm_base_url,
            temperature=0,
        )

        role = await get_prompt("food_recognize_role") or "你是一个专业的食材识别助手。"
        template = await get_prompt("food_recognize") or FOOD_RECOGNIZE_TEMPLATE
        try:
            prompt = template.format(image_url=req.image_url)
        except (KeyError, IndexError, ValueError):
            logger.warning("Java 配置的 food_recognize 模板缺少占位符，回退本地模板")
            prompt = FOOD_RECOGNIZE_TEMPLATE.format(image_url=req.image_url)

        response = llm.invoke([
            SystemMessage(content=role),
            HumanMessage(content=prompt),
        ])

        text = response.content.strip()
        try:
            foods = json.loads(text)
        except json.JSONDecodeError:
            match = re.search(r'\[[\s\S]*\]', text)
            foods = json.loads(match.group()) if match else []

        foods_list = []
        for item in foods:
            if isinstance(item, dict):
                foods_list.append(FoodsItem(
                    name=item.get("name", "未知"),
                    confidence=float(item.get("confidence", 0.5)),
                    category=item.get("category", "other"),
                ))

        return FoodRecognizeResponse(foods=foods_list, raw_response=text, trace_id=trace_id)
    except Exception as e:
        logger.error("食材识别失败: %s", e, exc_info=True)
        return FoodRecognizeResponse(foods=[], trace_id=trace_id)


@router.post("/menu/generate", response_model=MenuGenerateResponse)
async def generate_menu(req: MenuGenerateRequest):
    """菜单生成: 根据食材和用餐人数生成一日三餐菜单"""
    trace_id = str(uuid.uuid4())
    try:
        from langchain_openai import ChatOpenAI
        from langchain_core.messages import HumanMessage, SystemMessage

        try:
            selected_foods = json.loads(req.selected_foods) if req.selected_foods else []
        except json.JSONDecodeError:
            selected_foods = []

        try:
            participants = json.loads(req.participants) if req.participants else []
        except json.JSONDecodeError:
            participants = []

        participant_count = len(participants) if participants else 1
        foods_text = ", ".join([f.get("name", "") for f in selected_foods]) if selected_foods else "根据可用食材"

        constraints = []
        if req.allergies:
            constraints.append(f"禁忌: {req.allergies}")
        if req.health_goals:
            constraints.append(f"健康目标: {req.health_goals}")
        if req.cuisine_pref:
            constraints.append(f"菜系偏好: {req.cuisine_pref}")
        if req.spice_level is not None:
            constraints.append(f"辣度: {req.spice_level}/5")

        # 菜数要求（CR-C / R-05）：注入既有 {constraints_text} 占位符，不新增占位符（F-18），
        # 因此 DB 模板（get_prompt）与本地模板两个来源同时生效；不依赖 prompt 的正确性，
        # 菜数硬保证由 Java DietRecommendationService.normalizeMenu 兜底。
        # Java AiGateway 传 dish_count = D(N)；旧调用方不传时按 participant_count 现算，行为一致。
        effective_dish_count = req.dish_count if req.dish_count is not None else 4 + max(0, participant_count - 3)
        constraints.append(
            f"菜数硬要求: 午餐和晚餐各需 {effective_dish_count} 道菜 + 1 道汤"
            f"（D(N)=4+max(0,N-3)，汤恒 1 且不计入菜数、只加菜不加汤；"
            f"汤标 type=soup、菜标 type=dish，汤恰好 1 道且排在最后）；早餐维持 1-2 道。"
        )
        constraints_text = "\n".join(constraints) if constraints else "无特殊限制"

        role = await get_prompt("menu_generate_role") or "你是一个专业营养师和厨师，擅长根据食材和健康目标设计食谱。"
        template = await get_prompt("menu_generate") or MENU_GENERATE_TEMPLATE
        try:
            prompt = template.format(
                participant_count=participant_count,
                date=req.date,
                foods_text=foods_text,
                constraints_text=constraints_text,
            )
        except (KeyError, IndexError, ValueError):
            logger.warning("Java 配置的 menu_generate 模板缺少占位符，回退本地模板")
            prompt = MENU_GENERATE_TEMPLATE.format(
                participant_count=participant_count,
                date=req.date,
                foods_text=foods_text,
                constraints_text=constraints_text,
            )

        llm = ChatOpenAI(
            # agnes-2.5-flash：非推理模型，菜单生成 11s（deepseek 推理型会占满 token 导致 content 为空）
            model="agnes-2.5-flash",
            api_key=settings.llm_api_key,
            base_url=settings.llm_base_url,
            temperature=0.7,
            max_tokens=3000,
        )

        response = llm.invoke([
            SystemMessage(content=role),
            HumanMessage(content=prompt),
        ])

        text = response.content.strip()
        match = re.search(r'\{[\s\S]*\}', text)
        menu_json = match.group() if match else '{"meals": []}'

        return MenuGenerateResponse(menu_json=menu_json, trace_id=trace_id)
    except Exception as e:
        logger.error("菜单生成失败: %s", e, exc_info=True)
        return MenuGenerateResponse(menu_json='{"meals": []}', trace_id=trace_id)