Large Language Models (LLMs) that rely solely on parametric memory learned through training have demonstrated strong performance in biomedical question-answering, but their tendency to hallucinate facts and the difficulty of adjusting and adding to the learned knowledge limit their usefulness in clinical settings. Retrieval-Augmented Generation (RAG) has emerged as a promising solution by adding a non-parametric source of memory and grounding LLM outputs to reputable external sources. This paper aims to survey the evolution of RAG methodologies and highlight the most recent developments and shifts in paradigms like Agentic RAG. We highlight current state-of-the-art biomedical RAG systems, the latest evaluation benchmarks, and empirical findings regarding best fine-tuning practices. We also examine technical including lost-in-the-middle effects and scaling behaviour. Ethical concerns surrounding privacy and bias are brought to attention alongside research gaps. Finally, we discuss future directions of RAG in the biomedical field, including integration into Clinical Decision Support systems (CDS).
Eason Ni (2026) studied this question.