c/backend/app/routes/qa_routes.py

from flask import Blueprint, current_app, request
from flask_jwt_extended import jwt_required

from app.rag.llm_client import LLMConfigManager, ask_dify, ask_fastgpt
from app.rag.local_retriever import LocalKnowledgeRetriever
from app.utils.auth import admin_required, current_user
from app.utils.response import ok


qa_bp = Blueprint("qa", __name__)


def _retriever() -> LocalKnowledgeRetriever:
    return LocalKnowledgeRetriever(current_app.config["LOCAL_KB_PATH"])


def _llm_config() -> dict:
    manager = LLMConfigManager(current_app.config["LLM_CONFIG_PATH"])
    return manager.load()


def _default_provider(config: dict) -> str:
    provider = (config.get("active_provider") or "fastgpt").strip().lower() if isinstance(config, dict) else "fastgpt"
    return provider if provider in {"fastgpt", "dify"} else "fastgpt"


def _context_text(hits: list) -> str:
    chunks = []
    for idx, item in enumerate(hits, start=1):
        chunks.append(f"[{idx}] 问: {item['question']}\n答: {item['answer']}")
    return "\n\n".join(chunks)


def _local_answer(hits: list) -> str:
    if not hits:
        return "本地知识库暂未检索到高相关内容，建议补充问题细节。"
    top = hits[0]
    return f"根据本地知识库：{top['answer']}"


def _advice_template(advice_type: str) -> str:
    templates = {
        "gain_muscle": "增肌建议：每天蛋白质建议 1.6-2.2g/kg，训练后 30 分钟内补充蛋白+碳水，优先鸡胸肉/牛肉/鸡蛋/燕麦。",
        "lose_fat": "减脂建议：保持 10%-20% 轻热量缺口，优先高蛋白高纤维食物，减少高糖饮料和深夜加餐。",
        "keto": "生酮建议：控制碳水通常小于 50g/天，优先健康脂肪和适量蛋白，并关注电解质补充。",
        "nutritionist": "营养师建议：结合体重、体脂、活动量和目标制定周计划，每 7 天复盘体重和围度变化。",
        "general": "通用建议：保证规律作息、三餐结构稳定、适量运动和充足饮水。",
    }
    return templates.get(advice_type, templates["general"])


def _ask_by_provider(provider: str, config: dict, question: str, context: str, user_id: str, username: str):
    if provider == "dify":
        return ask_dify(config, question, context=context, user=username)
    return ask_fastgpt(config, question, context=context, custom_uid=user_id)


@qa_bp.get("/kb/search")
@jwt_required(optional=True)
def search_kb():
    query = (request.args.get("query") or "").strip()
    top_k = min(max(int(request.args.get("top_k", 5) or 5), 1), 10)

    retriever = _retriever()
    hits = retriever.search(query, top_k=top_k)
    return ok({"items": hits})


@qa_bp.post("/ask")
@jwt_required()
def ask_nutrition_qa():
    payload = request.get_json(silent=True) or {}
    question = (payload.get("question") or "").strip()
    if not question:
        return ok({"answer": "请输入问题", "items": []}, "empty_question")

    mode = (payload.get("mode") or "auto").strip().lower()

    llm_cfg = _llm_config()
    provider = (payload.get("provider") or _default_provider(llm_cfg)).strip().lower()
    if provider not in {"fastgpt", "dify"}:
        provider = _default_provider(llm_cfg)

    retriever = _retriever()
    hits = retriever.search(question, top_k=3)
    context = _context_text(hits)

    local_answer = _local_answer(hits)
    if mode == "local":
        return ok({"answer": local_answer, "provider": "local_kb", "items": hits})

    user = current_user()
    llm_result = _ask_by_provider(
        provider,
        llm_cfg,
        question,
        context,
        str(user.id) if user else "",
        user.username if user else "anonymous",
    )

    if mode == "llm":
        if llm_result.get("ok"):
            return ok({"answer": llm_result.get("answer"), "provider": provider, "items": hits})
        return ok(
            {
                "answer": f"LLM 调用失败，已回退本地知识库。{local_answer}",
                "provider": "local_kb_fallback",
                "items": hits,
                "llm_error": llm_result,
            }
        )

    if llm_result.get("ok"):
        return ok({"answer": llm_result.get("answer"), "provider": provider, "items": hits})

    return ok(
        {
            "answer": local_answer,
            "provider": "local_kb",
            "items": hits,
            "llm_error": llm_result,
        }
    )


@qa_bp.post("/advice")
@jwt_required()
def advice():
    payload = request.get_json(silent=True) or {}
    advice_type = (payload.get("advice_type") or "general").strip().lower()
    question = (payload.get("question") or "请根据我的状态给建议").strip()

    llm_cfg = _llm_config()
    provider = (payload.get("provider") or _default_provider(llm_cfg)).strip().lower()
    if provider not in {"fastgpt", "dify"}:
        provider = _default_provider(llm_cfg)

    retriever = _retriever()
    hits = retriever.search(f"{advice_type} {question}", top_k=3)
    context = _context_text(hits)
    base_text = _advice_template(advice_type)

    user = current_user()
    llm_result = _ask_by_provider(
        provider,
        llm_cfg,
        f"请给出{advice_type}方向的个性化饮食建议。用户补充: {question}",
        f"{base_text}\n\n{context}",
        str(user.id) if user else "",
        user.username if user else "anonymous",
    )

    if llm_result.get("ok"):
        answer = llm_result.get("answer")
        source = provider
    else:
        answer = f"{base_text}\n\n（提示：LLM 暂不可用，当前为本地模板建议）"
        source = "local_template"

    return ok(
        {
            "advice_type": advice_type,
            "answer": answer,
            "provider": source,
            "knowledge_hits": hits,
            "llm_status": llm_result,
        }
    )


@qa_bp.post("/kb/reload")
@admin_required
def reload_kb():
    retriever = _retriever()
    retriever.reload()
    return ok({"count": len(retriever.documents)}, "知识库已重载")