Hello, it looks like a scenario where RAG handles multi-turn dialogues.

A preliminary idea: combine the previous rounds of dialogue with LLM to rewrite the current question, making the user's question clearer.