hunjing/backend/app/services/preferences.py

"""设置服务：模型网关状态聚合与生成偏好。

模型网关状态从 config/models.yaml 派生，provider 密钥按 settings 是否配置脱敏为
configured 布尔（绝不暴露值）；生成偏好为本地单用户单行，无则返回默认值。
"""

from app.ai_provider.config.loader import load_models_config
from app.ai_provider.config.models import ModelConfig
from app.core.config import settings
from app.core.time import utc_now
from app.db.session import AsyncSessionLocal
from app.models.governance import GenerationPreferences
from app.repositories import preferences as preferences_repo
from app.schemas.preferences import (
    GenerationPreferencesResponse,
    LogicalModel,
    ModelEndpointStatus,
    ModelGatewayStatus,
    ProviderKeyStatus,
    UpdateGenerationPreferencesRequest,
)
from app.services.ai_provider_admin import get_endpoint_overrides_cache
from app.services.serializers import generation_preferences_to_response

# provider → 持有其 API Key 的环境变量名；未登记的 provider 回退 <PROVIDER>_API_KEY。
_PROVIDER_KEY_ENV: dict[str, str] = {
    "openai": "OPENAI_API_KEY",
    "deepseek": "DEEPSEEK_API_KEY",
    "anthropic": "ANTHROPIC_API_KEY",
    "google": "GOOGLE_API_KEY",
    "minimax": "MINIMAX_API_KEY",
    "zhipu": "ZHIPU_API_KEY",
    "tencentcloud": "TENCENTCLOUD_API_KEY",
    "volcengine": "VOLCENGINE_ARK_API_KEY",
    "bailian": "BAILIAN_API_KEY",
}


def provider_configured(provider: str | None) -> bool:
    """provider 的 API Key 是否已在 settings 配置（仅返回布尔，不暴露值）。"""
    if not provider:
        return False
    env_name = _PROVIDER_KEY_ENV.get(provider, f"{provider.upper()}_API_KEY")
    return bool(getattr(settings, env_name.lower(), None))


def _endpoint_status(
    endpoint, model_name: str, provider: str, overrides_cache
) -> ModelEndpointStatus:
    ov = overrides_cache.get((model_name, endpoint.name))
    return ModelEndpointStatus(
        name=endpoint.name,
        provider=provider,
        status=(ov.get("status") or endpoint.status) if ov else endpoint.status,
        weight=endpoint.weight,
        priority=endpoint.priority,
        is_overridden=ov is not None,
        has_url_override=bool(ov and ov.get("base_url")),
        has_key_override=bool(ov and ov.get("api_key")),
        base_url_env=endpoint.base_url_env,
        api_key_env=endpoint.api_key_env,
    )


def _logical_model(cfg: ModelConfig, overrides_cache) -> LogicalModel:
    provider = cfg.provider or "unknown"
    endpoints = [
        _endpoint_status(ep, cfg.name, provider, overrides_cache) for ep in cfg.endpoints
    ] or None
    return LogicalModel(
        name=cfg.name,
        display_name=cfg.display_name or cfg.name,
        provider=provider,
        capabilities=[k for k, v in cfg.resolved_capabilities().items() if v],
        routing_strategy=cfg.routing_strategy,
        supports_streaming=bool(cfg.supports_streaming),
        supports_thinking=bool(cfg.supports_thinking),
        supports_structured_output=bool(cfg.supports_structured_output),
        context_window=cfg.context_window,
        max_output_tokens=cfg.max_output_tokens,
        endpoints=endpoints,
        is_default=True if cfg.name == settings.default_model_hint else None,
    )


def get_provider_gateway() -> ModelGatewayStatus:
    """从 config/models.yaml 派生模型网关总览（含 DB 覆盖标记，脱敏不回原值）。"""
    models_cfg = load_models_config()
    overrides_cache = get_endpoint_overrides_cache()
    providers_seen: dict[str, None] = {}
    logical_models: list[LogicalModel] = []
    for cfg in models_cfg:
        logical_models.append(_logical_model(cfg, overrides_cache))
        provider = cfg.provider or "unknown"
        providers_seen.setdefault(provider, None)
    providers = [
        ProviderKeyStatus(provider=name, configured=provider_configured(name))
        for name in providers_seen
    ]
    return ModelGatewayStatus(
        models=logical_models,
        providers=providers,
        default_model_hint=settings.default_model_hint,
    )


def _default_preferences() -> GenerationPreferences:
    return GenerationPreferences(
        owner_scope=settings.default_owner_scope,
        default_model_hint=settings.default_model_hint,
        temperature=0.7,
        top_p=1.0,
        max_output_tokens=4096,
        streaming=True,
    )


async def get_generation_preferences(
    owner_scope: str = settings.default_owner_scope,
) -> GenerationPreferencesResponse:
    async with AsyncSessionLocal() as session:
        prefs = await preferences_repo.get_preferences(session, owner_scope)
        if prefs is None:
            prefs = _default_preferences()
        return generation_preferences_to_response(prefs)


async def update_generation_preferences(
    payload: UpdateGenerationPreferencesRequest,
    owner_scope: str = settings.default_owner_scope,
) -> GenerationPreferencesResponse:
    async with AsyncSessionLocal() as session:
        prefs = await preferences_repo.get_preferences(session, owner_scope)
        if prefs is None:
            prefs = _default_preferences()
            for key, value in payload.model_dump(exclude_unset=True).items():
                setattr(prefs, key, value)
            prefs.updated_at = utc_now()
            session.add(prefs)
            try:
                await session.commit()
            except Exception:
                await session.rollback()
                prefs = await preferences_repo.get_preferences(session, owner_scope)
                if prefs is None:
                    raise
                for key, value in payload.model_dump(exclude_unset=True).items():
                    setattr(prefs, key, value)
                prefs.updated_at = utc_now()
                await session.commit()
        else:
            for key, value in payload.model_dump(exclude_unset=True).items():
                setattr(prefs, key, value)
            prefs.updated_at = utc_now()
            await session.commit()
        await session.refresh(prefs)
        return generation_preferences_to_response(prefs)