RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards