CiteGuard-RAG: A Validation-Centered AI System for Evidence-Grounded Question Answering
CiteGuard-RAG puts a validation step between retrieval and delivery to decide whether an answer should go out, be refused, or be regenerated.
The paper combines hybrid retrieval, citation-constrained generation, sentence-level grounding checks, and single-pass regeneration. It was tested on 400 questions across housing-law data, PrivacyQA, and CUAD. In the controlled setup, it reports 99.1% retrieval accuracy, 98.3% grounded-answer accuracy, 98.3% citation validity, and no validation-detected hallucinations. Under domain shift, citation validity held up better than evidence use, span alignment, and refusal calibration. ArXiv · AI/CL/LG's note
The paper combines hybrid retrieval, citation-constrained generation, sentence-level grounding checks, and single-pass regeneration. It was tested on 400 questions across housing-law data, PrivacyQA, and CUAD. In the controlled setup, it reports 99.1% retrieval accuracy, 98.3% grounded-answer accuracy, 98.3% citation validity, and no validation-detected hallucinations. Under domain shift, citation validity held up better than evidence use, span alignment, and refusal calibration. ArXiv · AI/CL/LG's note
score 4