Sparse, Dense, or Hybrid? Comparing Retrieval Strategies for Biomedical Question Answering with Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) has emerged as a dominant paradigm for grounding large language models in external knowledge, yet the choice of retrieval strategy remains underexplored in the biomedical domain. This study presents an empirical comparison of four retrieval strategies---BM25 (sparse), Contriever (general-purpose dense), MedCPT (domain-specific dense), and a reciprocal rank fusion hybrid combining BM25 with MedCPT---within a standardized RAG pipeline for biomedical question answering. Experiments are conducted on three established benchmarks: PubMedQA, MedQA, and BioASQ Task B. Evaluation spans retrieval quality (Recall@10, Recall@20, MRR@10), end-to-end QA accuracy, and answer faithfulness measured through the RAGAS metric. Results indicate that the hybrid strategy achieves the highest Recall@10 across all three datasets, reaching 0.761 on PubMedQA, 0.697 on MedQA, and 0.768 on BioASQ. The domain-specific MedCPT retriever consistently outperforms the general-purpose Contriever, while BM25 remains a competitive baseline that surpasses Contriever on two of three benchmarks. End-to-end QA accuracy follows a similar pattern, with the hybrid strategy yielding the best performance at 0.741 on PubMedQA and 0.613 on MedQA. Faithfulness analysis reveals that domain-specific retrieval reduces hallucination rates by providing more topically relevant context. These findings offer practical guidance for practitioners selecting retrieval strategies when deploying biomedical RAG applications.