We propose a framework that utilizes LLM to support RL models. This framework refines RL decisions based on real-world contexts and provides reasonable actions when RL agents make erroneous decisions.