Which of the following BEST reduces the risk of exposing sensitive data through the output of large language models (LLMs) in applications?