We propose a novel RAG framework in which a retriever and an LLM cooperate to exchange knowledge, and the retriever utilizes information of both earlier and later layers to accurately rank documents relevant to a question.