This paper introduces an approach for training o1-like RAG models that retrieve and reason over relevant information step by step before generating the final answer.