LLM max_tokens 1000→8000: reasoning-модели тратили весь лимит на размышления и возвращали пустой ответ

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.env

@@ -23,5 +23,5 @@ LLM_BASE_URL=
 LLM_API_KEY=
 LLM_MODEL=
 LLM_TIMEOUT=120
-LLM_MAX_TOKENS=1000
+LLM_MAX_TOKENS=8000
 LLM_TEMPERATURE=0.3

api/config.py

@@ -20,7 +20,8 @@ class Settings(BaseSettings):
     LLM_API_KEY: str = os.getenv("LLM_API_KEY", "")
     LLM_MODEL: str = os.getenv("LLM_MODEL", "")
     LLM_TIMEOUT: int = int(os.getenv("LLM_TIMEOUT", "120"))
-    LLM_MAX_TOKENS: int = int(os.getenv("LLM_MAX_TOKENS", "1000"))
+    # reasoning-модели тратят max_tokens и на размышления — 1000 им мало
+    LLM_MAX_TOKENS: int = int(os.getenv("LLM_MAX_TOKENS", "8000"))
     LLM_TEMPERATURE: float = float(os.getenv("LLM_TEMPERATURE", "0.3"))
 
     @property

api/lib/llm.py

@@ -36,7 +36,8 @@ class LLMConfig(BaseModel):
     # before the user has picked one.
     model: str = ""
     timeout: int = 120
-    max_tokens: int = 1000
+    # у reasoning-моделей сюда входят и токены размышлений
+    max_tokens: int = 8000
     temperature: float = 0.3
     extra_headers: dict[str, str] = Field(default_factory=dict)
 

api/route/pipeline.py

@@ -18,7 +18,7 @@ class LLMOverride(BaseModel):
     api_key: str = ""
     model: str
     timeout: int = 120
-    max_tokens: int = 1000
+    max_tokens: int = 8000
     temperature: float = 0.3
     extra_headers: dict[str, str] = {}
     system_prompt: str | None = None