Policy-Guided RAG is a novel Retrieval-Augmented Generation (RAG) framework that enables organizations to enforce hidden business rules, safety policies, and compliance constraints during the retrieval process without exposing those policies to the underlying Large Language Model (LLM). The system uses an asymmetric visibility architecture where policy guidance influences document ranking through query augmentation and reranking mechanisms but is filtered out before generation, ensuring zero policy leakage. Evaluated on credit card recommendation tasks, the approach achieved up to 13.3% improvement in Top-5 retrieval accuracy on curated datasets and 9.3% improvement at scale, with particularly strong performance on complex queries. The solution demonstrates how controllable, policy-aware retrieval can improve recommendation quality while maintaining confidentiality, making it valuable for enterprise applications requiring regulatory compliance, business logic enforcement, and AI safety.
Jainil Gosalia (Tue,) studied this question.