Embedding Domain Knowledge for Large Language Models via Reinforcement Learning from Augmented Gen