An Agentic Retrieval-Augmented Generation Framework for Reliable Biomedical Evidence Summarization

Abstract

The rapid expansion of biomedical literature has made it difficult for researchers, clinicians, and healthcare professionals to extract reliable and evidence-supported information from large collections of scientific documents. Although modern language models can generate summaries efficiently, they may produce unsupported, incomplete, or factually incorrect statements, which limits their direct use in biomedical and clinical decision-support contexts. To address this limitation, this paper presents BioEvidence-RAG, an agentic retrieval-augmented generation framework for trustworthy biomedical literature summarization. The proposed framework integrates document retrieval, evidence selection, controlled summary generation, fact verification, citation mapping, and reliability evaluation within a unified multi-agent pipeline. Each agent performs a specific task and passes verified information to the next stage, ensuring that generated summary statements are grounded in retrieved biomedical evidence. The system operates on locally indexed biomedical corpora such as PubMed and BioASQ, supporting offline, secure, and reproducible execution without dependency on cloud-based services. Experimental evaluation shows strong performance across both conventional summarization metrics and faithfulness-oriented measures. The framework achieves factual accuracy above 96%, evidence coverage above 85%, and keeps unsupported content below 5%. Compared with standalone language model summarization, the retrieval-augmented and verification-based design reduces hallucination and improves context handling in multi-document biomedical summarization. Overall, the proposed framework improves trust, transparency, traceability, and reliability in biomedical text summarization and is suitable for academic research environments where evidence-grounded outputs are essential.