Taxonomy-Guided Query Rewriting for Indonesian Hemodialysis RAG

Abstract

Retrieval-Augmented Generation (RAG) for medical question answering requires not only topically relevant documents, but context that directly addresses the user's question. This study investigates how a domain-specific taxonomy can improve retrieval quality in an Indonesian hemodialysis patienteducation system. We build a nephrology taxonomy of 350 concepts and 1,841 synonym entries, and use it to reformulate queries through expansion and rewriting. We evaluate eight scenarios using BGE-M3 for retrieval and compare feature-based with cross-encoder reranking on 71 medically validated questions. Results show that query rewriting combined with cross-encoder reranking achieves the best performance: Faithfulness 0.9019, Answer Relevancy 0.9160, and Context Precision 0.8784. The taxonomy effectively standardizes term variants and supports query reformulation, but direct taxonomy scoring does not surpass the baseline. These findings suggest that taxonomy is most valuable at the query stage, with the cross-encoder handling final relevance judgment.