A Related Line of Research RegTutor: Self-correcting Agentic Retrieval-Augmented Generation for Medical-Device Compliance Training

Abstract

Medical-device compliance training requires feedback grounded in authoritative regulations. However, existing retrieval-augmented generation approaches often lack explicit evidence assessment and post-generation validation, weakening citation traceability in high-stakes regulatory tutoring. To address this limitation, we propose RegTutor, a self-correcting agentic retrieval-augmented generation framework for medical-device compliance training. Specifically, RegTutor retrieves regulatory evidence from 21 CFR Parts 803, 807, and 820, as well as FDA guidance, evaluates evidence sufficiency, rewrites inadequate queries, generates tutoring feedback, and validates citation format and claim-evidence alignment before finalizing its response. Extensive results showed that RegTutor achieved 100.0% answer correctness, citation accuracy, and helpfulness on a predefined 30-item controlled evaluation. In a balanced 300-item operational rerun, all RegTutor runs completed successfully, and citation accuracy reached 99.0%. On a separately frozen 60-item manually reviewed blind evaluation, RegTutor achieved 100.0% answer correctness, citation accuracy, and helpfulness among 59 completed runs. A secondary replication with GLM-4.7 further demonstrated 100.0% answer correctness and 98.3% citation accuracy. These findings demonstrate the value of self-correction and citation validation for traceable regulatory tutoring