A Generative AI Engineer is developing a RAG application for customer support. The application uses product manuals, chat logs, and user-generated content as part of the knowledge base. During a security audit, concerns are raised about the possibility of the model reproducing abusive or harmful language from chat logs during inference. Which approach is MOST appropriate to mitigate this risk without significantly degrading model performance or coverage?