Update main.py

main.py

@@ -11,18 +11,34 @@ app = FastAPI(title="CygnisAI Studio API")
 HF_TOKEN = os.environ.get("HF_TOKEN")
 CYGNIS_API_KEY = os.environ.get("CYGNIS_API_KEY", "cgn_live_stable_demo_api_key_012345")
 
+# Mapping vers des modèles DISPONIBLES sur le routeur gratuit
 MODELS = {
-    "google/gemma-3-27b-it": "google/gemma-2-27b-it", 
+    # Gemma 2 9B (Google) - Très rapide et dispo
+    "google/gemma-3-27b-it": "google/gemma-2-9b-it", 
+    
+    # Llama 3.1 70B (Meta) - Puissant
     "openai/gpt-oss-120b": "meta-llama/Meta-Llama-3.1-70B-Instruct", 
-    "Qwen/Qwen3-VL-8B-Thinking": "Qwen/Qwen2.5-72B-Instruct",
-    "XiaomiMiMo/MiMo-V2-Flash": "Xiaomi/MIMO", 
-    "deepseek-ai/DeepSeek-V3.2": "deepseek-ai/DeepSeek-V3", 
-    "meta-llama/Llama-4-Scout-17B-16E-Instruct": "meta-llama/Meta-Llama-3.1-8B-Instruct", 
-    "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": "nvidia/Llama-3.1-Nemotron-70B-Instruct-HF", 
-    "default": "meta-llama/Meta-Llama-3-8B-Instruct"
+    
+    # Qwen 2.5 7B (Alibaba) - Excellent généraliste
+    "Qwen/Qwen3-VL-8B-Thinking": "Qwen/Qwen2.5-7B-Instruct",
+    
+    # Phi 3.5 (Microsoft) - Léger
+    "XiaomiMiMo/MiMo-V2-Flash": "microsoft/Phi-3.5-mini-instruct", 
+    
+    # DeepSeek R1 (Distill Llama 8B) - Raisonnement
+    "deepseek-ai/DeepSeek-V3.2": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", 
+    
+    # Llama 3.2 3B (Meta) - Ultra rapide
+    "meta-llama/Llama-4-Scout-17B-16E-Instruct": "meta-llama/Llama-3.2-3B-Instruct", 
+    
+    # Llama 3.1 8B (Meta) - Standard
+    "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": "meta-llama/Meta-Llama-3.1-8B-Instruct", 
+    
+    # Défaut : Llama 3.1 8B (Le plus stable sur le routeur)
+    "default": "meta-llama/Meta-Llama-3.1-8B-Instruct"
 }
 
-# NOUVELLE URL DE BASE UNIQUE
+# URL de base du routeur d'inférence HF
 HF_ROUTER_BASE = "https://router.huggingface.co/hf-inference/models"
 
 class ChatRequest(BaseModel):
@@ -78,7 +94,6 @@ async def ask_model(req: ChatRequest, authorized: bool = Depends(verify_api_key)
 
     try:
         # 1. Tentative via endpoint Chat (OpenAI compatible)
-        # URL: https://router.huggingface.co/hf-inference/models/{model_id}/v1/chat/completions
         hf_chat_url = f"{HF_ROUTER_BASE}/{model_id}/v1/chat/completions"
         
         payload_chat = {
@@ -92,14 +107,16 @@ async def ask_model(req: ChatRequest, authorized: bool = Depends(verify_api_key)
         print(f"🚀 Calling HF Chat API: {hf_chat_url}")
         response = requests.post(hf_chat_url, headers=headers, json=payload_chat)
         
-        # 2. Fallback via endpoint Inference Standard (si Chat échoue avec 404 ou 405)
+        # 2. Fallback via endpoint Inference Standard
         if response.status_code in [404, 405]:
              print(f"🔄 Fallback to standard inference API (Status {response.status_code})")
-             
-             # URL: https://router.huggingface.co/hf-inference/models/{model_id}
              api_url = f"{HF_ROUTER_BASE}/{model_id}"
              
-             prompt_str = f"System: {req.system_prompt}\nUser: {req.question}\nAssistant:" if req.system_prompt else f"User: {req.question}\nAssistant:"
+             # Formatage simple pour le fallback
+             prompt_str = ""
+             for msg in messages:
+                 prompt_str += f"{msg['role']}: {msg['content']}\n"
+             prompt_str += "assistant:"
              
              payload_standard = {
                  "inputs": prompt_str,