Answers are grounded in this paper and cite their sources.
Retrieval-augmented generation (RAG) has advanced biomedical question answering (QA), but existing systems are limited by single-pass retrieval and sensitivity to lexical variability in domain-specific terminology. This study proposes an adaptive RAG framework that integrates logically structured query expansion with multi-hop retrieval and reasoning to address these challenges. Queries are reformulated using a propositional conjunctive normal form to capture semantic variants while preserving logical consistency. A classification module distinguishes between single-hop and multi-hop questions, enabling adaptive processing. For complex queries, the framework performs query decomposition and generates intermediate supporting information. Retrieved documents are further refined through sentence-level segmentation and embedding-based re-ranking to identify the most relevant evidence. The approach is evaluated on the BioASQ Task B benchmark using standard metrics and RAGAS-based evaluation. Results show that the full framework achieved F1 ROUGE-2 and ROUGE-SU4 scores of 0.161 and 0.162, respectively, compared with 0.157 and 0.157 for the Naive RAG baseline, indicating modest improvements in ideal-answer quality. The framework was particularly useful for complex multi-hop questions, while simpler configurations remained competitive on several exact-answer metrics. Expert evaluation confirms that the generated answers are clear, concise, and clinically meaningful. Overall, the framework offers a practical and interpretable solution for improving biomedical QA systems.
No takes yet. Share an insight, caveat, or question.
Tran et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: