Improve HF ranking compatibility by enabling chat endpoint fallback.

Allow scoring requests to try router chat completions after generate/completions so ranking can succeed on providers exposing chat-only interfaces.

Co-authored-by: Cursor <cursoragent@cursor.com>

llm_service.py

@@ -18,9 +18,10 @@ logger = logging.getLogger(__name__)
 # Router statuses that should trigger trying the next endpoint format.
 _ROUTER_FALLBACK_STATUSES = (400, 404, 405, 415, 422)
 _DEFAULT_FALLBACK_MODELS = (
-    "Qwen/Qwen2.5-32B-Instruct",
-    "meta-llama/Llama-3.1-70B-Instruct",
     "Qwen/Qwen2.5-7B-Instruct",
+    "mistralai/Mistral-7B-Instruct-v0.3",
+    "google/gemma-2-9b-it",
+    "meta-llama/Llama-3.1-8B-Instruct",
 )
 _BIOMEDICAL_SYNONYMS = {
     "kras": ["k-ras", "kras g12c", "kras g12d", "krasi"],
@@ -582,7 +583,7 @@ REASONING: [brief, natural explanation of why this score was assigned]"""
         """Score using Hugging Face Inference API (legacy method)"""
         try:
             response, _ = self._call_hf_router_api(
-                prompt, max_tokens=10, temperature=0.1, top_p=1.0, timeout=30, allow_chat_fallback=False
+                prompt, max_tokens=10, temperature=0.1, top_p=1.0, timeout=30, allow_chat_fallback=True
             )
             response.raise_for_status()
             generated_text = self._extract_text_from_hf_response(response.json())
@@ -599,7 +600,7 @@ REASONING: [brief, natural explanation of why this score was assigned]"""
                 return 0.5, "Hugging Face token missing. Add HUGGINGFACE_API_TOKEN in Space Secrets."
 
             response, used_model = self._call_hf_router_api(
-                prompt, max_tokens=200, timeout=60, allow_chat_fallback=False
+                prompt, max_tokens=200, timeout=60, allow_chat_fallback=True
             )
             
             if not response.ok:
@@ -615,7 +616,10 @@ REASONING: [brief, natural explanation of why this score was assigned]"""
                 if response.status_code == 429:
                     return 0.5, "Hugging Face rate limit reached (429). Please retry shortly."
                 if response.status_code in (400, 404, 422):
-                    return 0.5, f"Model/endpoint mismatch ({response.status_code}). Verify DEEPSEEK_MODEL is available for HF router."
+                    return 0.5, (
+                        f"Model/endpoint mismatch ({response.status_code}). "
+                        "Set DEEPSEEK_MODEL or HF_FALLBACK_MODELS to router-supported models."
+                    )
 
                 return 0.5, f"Hugging Face API error {response.status_code}: {error_preview[:140]}"