Running summarizing tools on a new article

Abstract: Biomedical question answering systems increasingly rely on retrieval augmented generation, but their outputs remain vulnerable to hallucinations when relevant evidence is absent, weakly ranked, or not available in the candidate list. We present a practical biomedical retrieval pipeline that improves evidence availability for generation while preserving transparency and bounded cost. The method combines readable lexical query variants over a local PubMed index with supervised keyword selection, bounded retrieval enrichment, weighted reciprocal rank fusion, and compact neural reranking. We evaluate five system variants, from a question-only lexical baseline to a full enriched and reranked pipeline, using retrieval metrics at 20, 50, and 100 documents to measure how much relevant evidence becomes visible to a generator. The results show that learned-keyword lexical retrieval improves over the raw question baseline, retrieval enrichment improves broader evidence coverage, direct reranking yields strong early gains, and the full enriched and reranked configuration gives the strongest overall retrieval results. These findings suggest lower hallucination risk through better evidence access rather than providing direct end-to-end proof on generated answers. The study positions biomedical retrieval as a useful intelligent-systems component for healthcare-oriented scientific literature search and safer generative AI workflows.