fix

backend/env/sql_env.py

@@ -67,15 +67,17 @@ class RewardInfo(BaseModel):
 
 # ─── LLM Client ──────────────────────────────────────────────────
 
+API_BASE_URL = os.environ.get("API_BASE_URL", "https://router.huggingface.co/v1")
+_MODEL = os.environ.get("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
+HF_TOKEN = os.environ.get("HF_TOKEN")  # no default — must be set explicitly
+
+
 def _make_client() -> AsyncOpenAI:
     return AsyncOpenAI(
-        api_key=os.environ.get("HF_TOKEN", ""),
-        base_url=os.environ.get("API_BASE_URL", "https://router.huggingface.co/v1"),
+        api_key=HF_TOKEN,
+        base_url=API_BASE_URL,
     )
 
-
-_MODEL = os.environ.get("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
-
 BASE_SYSTEM_PROMPT = """You are a SQL expert. Given a natural language question and a SQLite database schema, write a correct SQL query.
 
 Rules:

backend/gepa/optimizer.py

@@ -24,7 +24,9 @@ from pydantic import BaseModel
 _DATA_DIR = Path(os.environ.get("DATA_DIR", Path(__file__).parent.parent / "data"))
 GEPA_PATH = _DATA_DIR / "gepa_prompt.json"
 
-_MODEL = os.environ.get("MODEL_NAME", "gpt-4o-mini")
+_API_BASE_URL = os.environ.get("API_BASE_URL", "https://router.huggingface.co/v1")
+_MODEL = os.environ.get("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
+_HF_TOKEN = os.environ.get("HF_TOKEN")  # no default — must be set explicitly
 
 # How many queries between each GEPA optimization cycle.
 # Override with the GEPA_OPTIMIZE_EVERY environment variable.
@@ -62,8 +64,8 @@ class Candidate(BaseModel):
 
 def _make_client() -> AsyncOpenAI:
     return AsyncOpenAI(
-        api_key=os.environ.get("HF_TOKEN", ""),
-        base_url=os.environ.get("API_BASE_URL", "https://router.huggingface.co/v1"),
+        api_key=_HF_TOKEN,
+        base_url=_API_BASE_URL,
     )
 
 

backend/main.py

@@ -87,7 +87,7 @@ async def startup_event():
         print(f"Warning: database seed failed: {e}")
 
     # Log LLM config so it's visible in container logs
-    token = os.environ.get("HF_TOKEN", "")
+    token = os.environ.get("HF_TOKEN")  # no default
     api_base = os.environ.get("API_BASE_URL", "https://router.huggingface.co/v1")
     model = os.environ.get("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
     token_status = f"set ({len(token)} chars)" if token else "NOT SET"