Update app.py

app.py

@@ -204,6 +204,19 @@ def get_best_response(responses):
     best_response_index = total_similarities.argmax()
     return responses[best_response_index]
 
+async def generate_model_response(model, inputs, max_tokens=2048):
+    try:
+        response = model(inputs)
+        text = remove_duplicates(response['choices'][0]['text'])
+        
+        if len(text.split()) > max_tokens:
+            parts = [text[i:i + max_tokens] for i in range(0, len(text), max_tokens)]
+            return ' '.join(parts)
+        
+        return text
+    except Exception as e:
+        return ""
+
 async def process_message(message):
     inputs = normalize_input(message)
     with ThreadPoolExecutor() as executor: