arxiv Reinforcement Learning for Optimizing RAG for Domain Chatbots