Apply RAG to pre-trained Large Language Models