How would you design a production RAG system to minimize hallucination and keep answers grounded in retrieved documents?