That is extremely hard, It require the model to have "knowledge" so it can decide if it know the answer or not which is not how current llm/gpt works