Update app.py

app.py

@@ -18,7 +18,7 @@ models = [
     {"repo_id": "Ffftdtd5dtft/gemma-2-27b-Q2_K-GGUF", "filename": "gemma-2-27b-q2_k.gguf"},
 ]
 
-# Cargar modelos en memoria
+# Cargar modelos en memoria solo una vez
 llms = [Llama.from_pretrained(repo_id=model['repo_id'], filename=model['filename']) for model in models]
 
 class ChatRequest(BaseModel):
@@ -29,7 +29,8 @@ class ChatRequest(BaseModel):
 
 def generate_chat_response(request, llm):
     try:
-        user_input = request.message
+        # Normalización del mensaje para manejo robusto
+        user_input = normalize_input(request.message)
         response = llm.create_chat_completion(
             messages=[{"role": "user", "content": user_input}],
             top_k=request.top_k,
@@ -41,6 +42,10 @@ def generate_chat_response(request, llm):
     except Exception as e:
         return {"response": f"Error: {str(e)}", "literal": user_input}
 
+def normalize_input(input_text):
+    # Implementar aquí cualquier lógica de normalización que sea necesaria
+    return input_text.strip()
+
 def select_best_response(responses, request):
     coherent_responses = filter_by_coherence([resp['response'] for resp in responses], request)
     best_response = filter_by_similarity(coherent_responses)
@@ -62,6 +67,9 @@ def filter_by_similarity(responses):
 
 @app.post("/generate_chat")
 async def generate_chat(request: ChatRequest):
+    if not request.message.strip():
+        raise HTTPException(status_code=400, detail="The message cannot be empty.")
+    
     with ThreadPoolExecutor(max_workers=None) as executor:
         futures = [executor.submit(generate_chat_response, request, llm) for llm in llms]
         responses = []