Update app/services/classifier.py

app/services/classifier.py

@@ -16,7 +16,14 @@ NSFW_WEIGHTS_PATH = MODELS_DIR / "model.safetensors"
 NSFW_THRESHOLD = 0.75
 
 LLM_MODELS = [
-    os.getenv("GAME_LLM_MODEL", "Qwen/Qwen2.5-1.5B-Instruct"),
+    x.strip()
+    for x in os.getenv("GAME_LLM_MODELS", os.getenv("GAME_LLM_MODEL", "Qwen/Qwen2.5-1.5B-Instruct")).split(",")
+    if x.strip()
+]
+LLM_PROVIDERS = [
+    x.strip()
+    for x in os.getenv("GAME_LLM_PROVIDERS", "auto,hf-inference").split(",")
+    if x.strip()
 ]
 LLM_MAX_CHARS = 3000
 
@@ -297,33 +304,64 @@ def _classify_game_with_llm(ocr_text: str, urls: list[str], suspected_game: bool
         {"role": "user", "content": json.dumps(user_payload, ensure_ascii=True)},
     ]
 
-    for model_name in LLM_MODELS:
-        try:
-            logger.info("llm try model=%s", model_name)
-            client = InferenceClient(model=model_name, token=token or None)
-            response = client.chat_completion(
-                messages=prompt_messages,
-                max_tokens=160,
-                temperature=0.1,
-                top_p=0.9,
-            )
-            content = ""
-            if response.choices:
-                content = response.choices[0].message.content or ""
-            parsed = _parse_llm_json(content)
-            if parsed is None:
-                logger.warning("llm invalid json model=%s content=%s", model_name, content[:300])
+    # Last candidate is empty => let HF pick a recommended chat model.
+    model_candidates = list(LLM_MODELS) + [""]
+
+    for provider_name in LLM_PROVIDERS:
+        for model_name in model_candidates:
+            model_label = model_name or "<recommended>"
+            try:
+                logger.info("llm try provider=%s model=%s", provider_name, model_label)
+                client = InferenceClient(
+                    model=model_name or None,
+                    provider=provider_name,
+                    token=token or None,
+                )
+                response = client.chat_completion(
+                    messages=prompt_messages,
+                    max_tokens=160,
+                    temperature=0.1,
+                    top_p=0.9,
+                )
+                content = ""
+                if response.choices:
+                    content = response.choices[0].message.content or ""
+                parsed = _parse_llm_json(content)
+                if parsed is None:
+                    logger.warning(
+                        "llm invalid json provider=%s model=%s content=%s",
+                        provider_name,
+                        model_label,
+                        content[:300],
+                    )
+                    continue
+                parsed["source"] = f"llm:{provider_name}:{model_label}"
+                return parsed
+            except Exception as exc:
+                if _is_model_not_supported_error(exc):
+                    logger.warning(
+                        "llm model/provider unsupported provider=%s model=%s",
+                        provider_name,
+                        model_label,
+                    )
+                    continue
+                logger.warning(
+                    "llm error provider=%s model=%s err=%s",
+                    provider_name,
+                    model_label,
+                    repr(exc),
+                )
                 continue
-            parsed["source"] = f"llm:{model_name}"
-            return parsed
-        except Exception as exc:
-            logger.warning("llm error model=%s err=%s", model_name, repr(exc))
-            continue
 
-    logger.warning("llm all models failed; fallback to heuristic")
+    logger.warning("llm all providers/models failed; fallback to heuristic")
     return None
 
 
+def _is_model_not_supported_error(exc: Exception) -> bool:
+    text = repr(exc).lower()
+    return "model_not_supported" in text or "not supported by any provider" in text
+
+
 def _parse_llm_json(content: str) -> dict[str, Any] | None:
     if not content:
         return None