Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning