life-echo/api/app/adapters/llm/deepseek_eval_judge.py

"""DeepSeek 评测台评审：模型别名解析 + ChatOpenAI 装配。"""

from __future__ import annotations

from langchain_openai import ChatOpenAI

from app.adapters.llm.openai_base_url import normalize_openai_compatible_base_url
from app.core.config import settings
from app.core.eval_judge_spec import EvalJudgeLlmSpec
from app.features.evaluation.constants import eval_cfg
from app.core.runtime_constants import llm_defaults


def resolve_deepseek_eval_judge_model(
    requested: str,
) -> tuple[str, dict | None, str | None]:
    """将模型名（含旧别名）规范为 V4 的 model id、extra_body 与 reasoning_effort。

    官方：deepseek-chat / deepseek-reasoner 将弃用，分别对应 v4-flash 非思考 / 思考。
    """
    m = (requested or "").strip()
    if m == "deepseek-chat":
        return (
            "deepseek-v4-flash",
            {"thinking": {"type": "disabled"}},
            None,
        )
    if m in (
        "deepseek-reasoner",
        "deepseek-r1",
    ):
        return (
            "deepseek-v4-flash",
            {"thinking": {"type": "enabled"}},
            "high",
        )
    if m == "deepseek-v4-pro":
        return ("deepseek-v4-pro", None, "high")
    if m in ("", "deepseek-v4-flash"):
        if eval_cfg.judge_deepseek_thinking_enabled:
            return (
                "deepseek-v4-flash",
                {"thinking": {"type": "enabled"}},
                "high",
            )
        return (
            "deepseek-v4-flash",
            {"thinking": {"type": "disabled"}},
            None,
        )
    if "flash" in m.lower() or m.startswith("deepseek-v4"):
        return (m, None, None)
    return (m, None, None)


def build_deepseek_eval_judge_spec(
    judge_model: str | None,
) -> EvalJudgeLlmSpec | None:
    """密钥缺失时返回 None。"""
    api_key = (settings.deepseek_api_key or "").strip()
    if not api_key:
        return None
    want = (judge_model or "").strip()
    base = normalize_openai_compatible_base_url(
        llm_defaults.deepseek_base_url,
        fallback="https://api.deepseek.com",
    )
    default_m = (eval_cfg.judge_deepseek_model or "deepseek-v4-flash").strip()
    combined = want or default_m
    model, extra, effort = resolve_deepseek_eval_judge_model(combined)
    ctx = int(eval_cfg.judge_deepseek_context_window_tokens)
    llm_kw: dict = {
        "api_key": api_key,
        "base_url": base,
        "model": model,
        "temperature": eval_cfg.judge_temperature,
    }
    if extra is not None:
        llm_kw["extra_body"] = extra
    if effort is not None:
        llm_kw["reasoning_effort"] = effort
    return EvalJudgeLlmSpec(
        llm=ChatOpenAI(**llm_kw),
        provider="deepseek",
        resolved_model=model,
        context_window_tokens=ctx,
    )