RLDBF: Enhancing LLMs Via Reinforcement Learning With DataBase FeedBack