How to reduce LLM hallucinations in production systems?