Volume 2 - Issue 3, May - June 2026

Large Language Models (LLMs) have demonstrated remarkable potential in healthcare knowledge processing; however, their susceptibility to hallucination and static knowledge boundaries significantly limits safe clinical deployment. This paper introduces MedRAG, a healthcare conversational assistant built on a multi-tier Retrieval-Augmented Generation (RAG) architecture that grounds every generated response in dynamically retrieved, authoritative medical literature. Three progressively sophisticated architectural configurations are designed, implemented, and systematically evaluated: Foundational RAG employing dense MedCPT retrieval, Optimised RAG combining hybrid BM25 and dense search with cross-encoder re-ranking, and Modular RAG integrating knowledge graph traversal with iterative faithfulness verification. Experiments across five standardised benchmarks—MedQA-USMLE, MedMCQA, PubMedQA, MMLU-Medical, and a private chronic-pain diagnostic corpus—demonstrate statistically significant performance improvements over standalone LLM baselines. The Modular RAG configuration paired with GPT-4-Turbo achieves 91.1% accuracy on MedQA-USMLE, equalling the Med-Gemini state-of-the-art, while the fully open-source variant with LLaMA-3-8B attains 79.7%, directly exceeding GPT-4-Turbo without retrieval (73.4%). Hallucination rates are reduced from 39.1% to 9.2% through RAGAS faithfulness verification. Ablation analysis identifies cross-encoder re-ranking (−6.0 pp) and knowledge graph retrieval (−4.9 pp) as the highest-impact architectural components. These findings demonstrate that retrieval quality is a more decisive performance determinant than generator model scale, enabling resource-constrained healthcare systems to achieve frontierlevel clinical AI performance.

Alur Taher Basha, Mr.P.Bharath Kumar, Dr.D.William Albert,"Deep Learning and Retrieval-Augmented Generation Methods for Autonomous Clinical Diagnosis and Decision Support" International Journal of Advanced Multidisciplinary Research and Educational Development, V2(3): Page(715-720) May-June 2026. ISSN: 3107-6513. www.ijamred.com. Published by Scientific and Academic Research Publishing.