<![CDATA[Legal question answering is essential for compliance, dispute resolution, and everyday HR decision-making, yet large language models may produce persuasive but incorrect legal statements when supporting evidence is incomplete. While Retrieval-Augmented Generation and graph-based retrieval can ground responses in statutes and structured relations, Vietnamese legal QA often lacks an explicit, automated quality-control step that scores an answer, decides whether it should be refined, and checks that citations are actually supported. In this paper, we propose a self-reflection mechanism that adds an iterative generate–evaluate–refine loop to a Graph-RAG pipeline for Vietnamese labor-law questions. Each draft is evaluated with a hybrid score that combines how closely the answer matches retrieved legal context with a model-derived confidence estimate, and the system iterates until it reaches a quality threshold or a stopping limit. On a Vietnamese Labor Law benchmark, the approach improves accuracy from 81.5% to 86.7% and reduces hallucination from 18.7% to 9.3%, with only a modest increase in end-to-end latency in typical use. We also examine component contributions and remaining failure cases, finding that pairing contextual alignment with confidence produces more stable answers than relying on a single signal. These results indicate that self-reflection can serve as a lightweight, deployment-friendly safety layer for high-stakes legal QA without requiring additional labeled data or model fine-tuning, and it can be adapted to other Vietnamese legal domains that demand transparent, article- and clause-level evidence.]]>
Copyrights © 2026
