feat(api): DeepSeek V4 Flash 默认、HTTP 错讯与多供应商分层

- 主链路默认 deepseek-v4-flash，DEEPSEEK_THINKING_ENABLED 对齐旧非思考 chat
- 评测台评审装配迁入 adapters/llm（deepseek_eval_judge、zhipu_eval_judge）与 eval_judge_spec
- 拆分 llm_http_openai_chat_errors 与 llm_errors（DeepSeek/智谱品牌与文档链），llm_call 支持 http_error_vendor
- EvalJudgeService 按 spec.provider 传入 allm_json_call；评测台前端文案改为 V4 Flash
- 更新 .env 示例与 staging/production 的 DEEPSEEK_MODEL；补充 openai/供应商错讯测试

Made-with: Cursor

api/app/adapters/llm/deepseek.py

@@ -18,8 +18,10 @@ class DeepSeekLLMProvider:
         self,
         api_key: str,
         base_url: str = "https://api.deepseek.com",
-        model: str = "deepseek-chat",
+        model: str = "deepseek-v4-flash",
         temperature: float = 0.7,
+        *,
+        extra_body: dict | None = None,
     ):
         self._default_model = model
         self._default_temperature = temperature
@@ -28,6 +30,8 @@ class DeepSeekLLMProvider:
             "model": model,
             "api_key": api_key,
         }
+        if extra_body:
+            kwargs["extra_body"] = extra_body
         if base_url:
             cleaned = base_url.rstrip("/")
             for suffix in ("/v1/chat/completions", "/v1"):

api/app/adapters/llm/deepseek_eval_judge.py

@@ -0,0 +1,85 @@
+"""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
+
+
+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 settings.eval_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 settings.llm_api_key or "").strip()
+    if not api_key:
+        return None
+    want = (judge_model or "").strip()
+    base = normalize_openai_compatible_base_url(
+        settings.deepseek_base_url,
+        fallback="https://api.deepseek.com",
+    )
+    default_m = (settings.eval_judge_deepseek_model or "deepseek-v4-flash").strip()
+    combined = want or default_m
+    model, extra, effort = resolve_deepseek_eval_judge_model(combined)
+    ctx = int(settings.eval_judge_deepseek_context_window_tokens)
+    llm_kw: dict = {
+        "api_key": api_key,
+        "base_url": base,
+        "model": model,
+        "temperature": settings.eval_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,
+    )

api/app/adapters/llm/openai_base_url.py

@@ -0,0 +1,9 @@
+"""OpenAI/Chat Completions 兼容基址规范化（多供应商共用）。"""
+
+
+def normalize_openai_compatible_base_url(raw: str, *, fallback: str) -> str:
+    base = (raw or "").strip().rstrip("/") or fallback
+    for suffix in ("/v1/chat/completions", "/v1"):
+        if base.endswith(suffix):
+            base = base[: -len(suffix)]
+    return base

api/app/adapters/llm/zhipu_eval_judge.py

@@ -0,0 +1,37 @@
+"""智谱 GLM 评测台评审：OpenAI 兼容端点 + 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
+
+
+def build_zhipu_eval_judge_spec(
+    judge_model: str | None,
+) -> EvalJudgeLlmSpec | None:
+    """密钥缺失时返回 None。"""
+    api_key = (settings.eval_judge_api_key or settings.zhipu_api_key or "").strip()
+    if not api_key:
+        return None
+    want = (judge_model or "").strip()
+    base = normalize_openai_compatible_base_url(
+        settings.eval_judge_base_url,
+        fallback="https://open.bigmodel.cn/api/paas/v4",
+    )
+    model = want or (settings.eval_judge_model or "glm-5")
+    ctx = int(settings.eval_judge_context_window_tokens)
+    llm_kw: dict = {
+        "api_key": api_key,
+        "base_url": base,
+        "model": model,
+        "temperature": settings.eval_judge_temperature,
+    }
+    return EvalJudgeLlmSpec(
+        llm=ChatOpenAI(**llm_kw),
+        provider="zhipu",
+        resolved_model=model,
+        context_window_tokens=ctx,
+    )