Profeto/src/llm/predict.py

"""预测服务:拼上下文 → 调 LLM → 存预测。"""
from __future__ import annotations

import functools
import hashlib
import logging
import time
from dataclasses import dataclass
from datetime import datetime, timezone
from pathlib import Path
from threading import Lock

from src.core.config import settings
from src.db.base import AsyncSessionLocal
from src.db.models import Match, Prediction
from src.db.unit_of_work import get_uow
from src.llm.context_builder import build_context
from src.llm.provider import LLMProvider, get_default_provider

logger = logging.getLogger(__name__)

_PROMPT_DIR = Path(__file__).resolve().parent / "prompts"

# ── LLM 响应缓存(match+provider+model+version → 结果) ──
_CACHE_TTL_SEC = 300  # 5 分钟
_cache: dict[str, tuple[float, PredictResult]] = {}
_cache_lock = Lock()


def _cache_key(match_id: int, provider: str, model: str, version: str, tpl_hash: str) -> str:
    """缓存键:含 prompt 模板内容 hash。

    仅用 version 做键不够 —— 编辑器里改动 `match_prediction_v1.md` 而版本号
    不变时,进程内缓存仍会返回旧模板产生的旧结果(见审查报告 P2-6)。
    把模板内容 hash 纳入键,模板一改缓存自动失效。
    """
    return f"{match_id}:{provider}:{model}:{version}:{tpl_hash[:12]}"


def _get_cached(match_id: int, provider: str, model: str, version: str, tpl_hash: str) -> PredictResult | None:
    key = _cache_key(match_id, provider, model, version, tpl_hash)
    with _cache_lock:
        if key in _cache:
            ts, result = _cache[key]
            if time.time() - ts < _CACHE_TTL_SEC:
                return result
            del _cache[key]
    return None


def _set_cached(match_id: int, provider: str, model: str, version: str, tpl_hash: str, result: PredictResult) -> None:
    key = _cache_key(match_id, provider, model, version, tpl_hash)
    with _cache_lock:
        _cache[key] = (time.time(), result)


def clear_prompt_cache() -> None:
    """清空 prompt 模板缓存(供开发/热更新时手动调用)。

    lru_cache 的模板缓存是进程级的,改完 .md 需要重启进程才能生效;
    提供显式清理入口,避免"改了模板却看不到变化"的困惑(见审查报告 P2-5)。
    """
    _load_prompt_template.cache_clear()
    logger.info("prompt 模板缓存已清空")


@functools.lru_cache(maxsize=8)
def _load_prompt_template(version: str = "v1") -> str:
    """缓存 prompt 模板(进程生命周期内每个版本只读一次)。"""
    path = _PROMPT_DIR / f"match_prediction_{version}.md"
    if not path.exists():
        raise FileNotFoundError(f"prompt 模板不存在: {path}")
    with open(path, encoding="utf-8") as f:
        return f.read()


def _prompt_template_hash(version: str) -> str:
    """prompt 模板内容 hash(用于缓存键,模板变更即失效)。"""
    return hashlib.sha256(_load_prompt_template(version).encode("utf-8")).hexdigest()


@dataclass
class PredictResult:
    prediction_id: int
    provider: str
    model: str
    prompt_version: str
    pred_home_goals: float | None
    pred_away_goals: float | None
    pred_1x2: str | None
    subjective_confidence: float | None
    reasoning: str | None
    context: str
    latency_ms: int | None
    raw: dict | None


async def predict_match(
    match_id: int,
    *,
    provider: LLMProvider | None = None,
    model: str | None = None,
    prompt_version: str | None = None,
    mode: str = "multi",
    use_cache: bool = True,
) -> "PredictResult | MultiPredictResult":
    """预测入口。mode=multi(默认)走多 agent;mode=single 走单次调用。

    Args:
        use_cache: 是否允许返回进程内缓存结果。回测必须传 False——
            缓存命中不会新建 prediction 行,调用方会对同一个 prediction_id
            反复 settle,把不同比赛的真实比分覆盖到同一条记录上。
    """
    if mode == "single":
        return await _predict_single(
            match_id,
            provider=provider,
            model=model,
            prompt_version=prompt_version,
            use_cache=use_cache,
        )
    from src.llm.agents.orchestrator import predict_match_multi

    return await predict_match_multi(match_id, provider=provider, version=(prompt_version or "v1").removeprefix("multi_"))


async def _predict_single(
    match_id: int,
    *,
    provider: LLMProvider | None = None,
    model: str | None = None,
    prompt_version: str | None = None,
    use_cache: bool = True,
) -> PredictResult:
    """单次调用路径(原有实现)。"""
    if provider is None:
        provider = get_default_provider()
    if model:
        provider.model = model
    version = prompt_version or "v1"
    tpl_hash = _prompt_template_hash(version)

    # 0. 查缓存(同 match+provider+model+version+模板hash 5 分钟内直接返)
    if use_cache:
        cached = _get_cached(match_id, settings.LLM_PROVIDER, provider.model, version, tpl_hash)
        if cached is not None:
            logger.debug("predict cache hit match=%s", match_id)
            return cached

    # 1. 拼上下文
    ctx = await build_context(match_id)

    # 1.5 计算快照元数据(用于可复现性)
    now = datetime.now(timezone.utc)
    match_kickoff_at = ctx.match_dt
    prediction_cutoff_at = ctx.match_dt  # 默认:比赛时间作为数据截止
    input_hash = hashlib.sha256(ctx.text.encode("utf-8")).hexdigest()

    # 2. 拼 prompt(指定版本)
    template = _load_prompt_template(version)
    user_prompt = template.replace("{{context}}", ctx.text)

    # 3. 调 LLM
    resp = await provider.chat(
        system="你是一个严谨的足球预测专家。只输出 JSON。",
        user=user_prompt,
        json_mode=True,
        temperature=0.3,
        max_tokens=800,
    )

    if resp.error:
        raise RuntimeError(f"LLM error: {resp.error}")

    parsed = resp.parsed or {}

    # 3.5 严格校验 LLM 输出
    from src.llm.validation import validate_prediction_output
    try:
        validated = validate_prediction_output(parsed)
    except Exception as e:
        raise RuntimeError(f"LLM 输出校验失败: {e}")

    # 4. 存预测(使用 UnitOfWork 统一事务)
    async with get_uow() as session:
        # 验证 match 存在
        m = await session.get(Match, match_id)
        if m is None:
            raise ValueError(f"match {match_id} not found")

        pred = Prediction(
            match_id=match_id,
            provider=settings.LLM_PROVIDER,
            model=provider.model,
            prompt_version=version,
            prompt_tokens=resp.prompt_tokens,
            completion_tokens=resp.completion_tokens,
            latency_ms=resp.latency_ms,
            pred_home_goals=validated.pred_home_goals,
            pred_away_goals=validated.pred_away_goals,
            pred_1x2=validated.pred_1x2,
            subjective_confidence=validated.subjective_confidence,
            reasoning=validated.reasoning,
            raw_response=resp.raw,
            status="success",
            match_kickoff_at=match_kickoff_at,
            prediction_cutoff_at=prediction_cutoff_at,
            prediction_created_at=now,
            input_hash=input_hash,
        )
        session.add(pred)
        await session.refresh(pred)

        result = PredictResult(
            prediction_id=pred.id,
            provider=pred.provider,
            model=pred.model,
            prompt_version=version,
            pred_home_goals=pred.pred_home_goals,
            pred_away_goals=pred.pred_away_goals,
            pred_1x2=pred.pred_1x2,
            subjective_confidence=pred.subjective_confidence,
            reasoning=pred.reasoning,
            context=ctx.text,
            latency_ms=resp.latency_ms,
            raw=resp.raw,
        )

        # 5. 写入缓存(仅当允许缓存时)
        if use_cache:
            _set_cached(match_id, settings.LLM_PROVIDER, provider.model, version, tpl_hash, result)
        return result