RAG to riches: synthetic data for training RAG agents