gmTouringMiniApp/files/归档/server/app/llm.py

from __future__ import annotations

import json
import os
from datetime import datetime, timedelta
from typing import Any, Optional

from pydantic import ValidationError

from .knowledge import KnowledgeBase
from .prompts import SYSTEM_PROMPT
from .schemas import GeneratedItinerary, Pace, TravelPreferences


class LLMError(RuntimeError):
    pass


class LLMTimeout(LLMError):
    pass


class LLMService:
    def __init__(self, knowledge: KnowledgeBase) -> None:
        self.knowledge = knowledge
        self.mode = os.getenv("LLM_MODE", "mock").lower()
        self.timeout = float(os.getenv("LLM_TIMEOUT_SECONDS", "45"))

    @property
    def configured(self) -> bool:
        if self.mode == "mock":
            return True
        return bool(os.getenv("OPENAI_API_KEY") and os.getenv("OPENAI_MODEL"))

    async def generate(
        self,
        preferences: TravelPreferences,
        candidates: list[dict[str, Any]],
        previous: Optional[GeneratedItinerary] = None,
        user_message: Optional[str] = None,
    ) -> GeneratedItinerary:
        if self.mode == "mock":
            return self._mock_generate(preferences, candidates, previous, user_message)
        return await self._real_generate(preferences, candidates, previous, user_message)

    async def _real_generate(
        self,
        preferences: TravelPreferences,
        candidates: list[dict[str, Any]],
        previous: Optional[GeneratedItinerary],
        user_message: Optional[str],
    ) -> GeneratedItinerary:
        if not self.configured:
            raise LLMError("真实模型尚未配置")
        try:
            from openai import APITimeoutError, AsyncOpenAI
        except ImportError as exc:
            raise LLMError("缺少 openai 依赖") from exc

        client = AsyncOpenAI(
            api_key=os.environ["OPENAI_API_KEY"],
            base_url=os.getenv("OPENAI_BASE_URL") or None,
            timeout=self.timeout,
        )
        payload = {
            "preferences": preferences.model_dump(by_alias=True),
            "candidates": candidates,
            "travelTimes": self.knowledge.travel_context(
                candidates, preferences.transport.value
            ),
            "previousItinerary": (
                previous.model_dump(by_alias=True) if previous else None
            ),
            "adjustmentRequest": user_message,
        }
        feedback = ""
        allowed_ids = {place["id"] for place in candidates}

        for attempt in range(2):
            try:
                response = await client.chat.completions.create(
                    model=os.environ["OPENAI_MODEL"],
                    temperature=0.2,
                    response_format={"type": "json_object"},
                    messages=[
                        {"role": "system", "content": SYSTEM_PROMPT},
                        {
                            "role": "user",
                            "content": json.dumps(payload, ensure_ascii=False)
                            + feedback,
                        },
                    ],
                )
            except APITimeoutError as exc:
                raise LLMTimeout("模型调用超时") from exc
            except Exception as exc:
                raise LLMError(f"模型调用失败：{exc}") from exc

            try:
                content = response.choices[0].message.content or "{}"
                itinerary = GeneratedItinerary.model_validate_json(content)
                unknown = {
                    item.place_id for item in itinerary.items
                } - allowed_ids
                if unknown:
                    raise ValueError(f"包含未知地点ID：{sorted(unknown)}")
                return itinerary
            except (ValidationError, ValueError, json.JSONDecodeError) as exc:
                if attempt == 1:
                    raise LLMError(f"模型输出无法通过结构校验：{exc}") from exc
                feedback = f"\n上次输出校验失败：{exc}。请重新输出完整合法 JSON。"

        raise LLMError("模型未返回有效结果")

    def _mock_generate(
        self,
        preferences: TravelPreferences,
        candidates: list[dict[str, Any]],
        previous: Optional[GeneratedItinerary],
        user_message: Optional[str],
    ) -> GeneratedItinerary:
        if not candidates:
            raise LLMError("没有符合条件的候选地点")

        max_items = {
            Pace.RELAXED: 3,
            Pace.MODERATE: 4,
            Pace.COMPACT: 5,
        }[preferences.pace]
        target = 240 if preferences.duration.value == "half_day" else 480
        start = datetime(2026, 1, 1, 9, 0)
        elapsed = 0
        items = []
        previous_place: Optional[dict[str, Any]] = None

        for place in candidates:
            if len(items) >= max_items:
                break
            transfer_value: Optional[int] = 0
            if previous_place:
                transfer_value = self.knowledge.travel_minutes(
                    previous_place["id"],
                    place["id"],
                    preferences.transport.value,
                )
            transfer_for_math = transfer_value or 0
            duration = int(place.get("recommendedMinutes", 75))
            if elapsed + transfer_for_math + duration > target + 30 and items:
                continue
            item_start = start + timedelta(minutes=transfer_for_math)
            end = item_start + timedelta(minutes=duration)
            items.append(
                {
                    "startTime": item_start.strftime("%H:%M"),
                    "endTime": end.strftime("%H:%M"),
                    "placeId": place["id"],
                    "placeName": place["name"],
                    "activity": place["summary"],
                    "reason": self._reason(preferences, place),
                    "transferFromPreviousMinutes": (
                        transfer_value if previous_place else 0
                    ),
                    "tips": place.get("tips", [])[:2],
                }
            )
            elapsed += transfer_for_math + duration
            start = end
            previous_place = place

        themes = "、".join(preferences.themes or preferences.interests[:1])
        adjustment = f"；已响应“{user_message}”" if user_message else ""
        pace_label = {
            Pace.RELAXED: "轻松",
            Pace.MODERATE: "适中",
            Pace.COMPACT: "紧凑",
        }[preferences.pace]
        return GeneratedItinerary.model_validate(
            {
                "title": f"光明区{themes or '精选'}{'半日' if preferences.duration.value == 'half_day' else '一日'}游",
                "summary": f"以科学、人文与都市自然为线索，按{pace_label}节奏安排{adjustment}。",
                "totalMinutes": max(elapsed, 1),
                "estimatedCostText": "费用以场馆、景区及实际交通信息为准",
                "items": items,
                "notes": [
                    "开放时间、预约和票价请在出行前通过官方渠道再次确认。",
                    "交通耗时为POC估算，请以出发时的实际导航为准。",
                ],
            }
        )

    @staticmethod
    def _reason(
        preferences: TravelPreferences, place: dict[str, Any]
    ) -> str:
        matches = list(
            (set(preferences.themes) & set(place.get("themes", [])))
            | (set(preferences.interests) & set(place.get("interests", [])))
        )
        return (
            f"符合你的{'、'.join(sorted(matches))}偏好"
            if matches
            else "作为光明区同路线备选，便于控制整体节奏"
        )