How would you reduce hallucinations in a medical chatbot?