Applications, Technical Characteristics, and Evaluation Approaches of Retrieval-Augmented Generation in Nursing: Scoping Review
Journal of Medical Internet Research ·
Background: Large language models (LLMs) show great potential for widespread application in the field of nursing; however, their limitations include hallucinations, insufficient domain knowledge, and delays in updating evidence. Retrieval-augmented generation (RAG), by combining external knowledge retrieval with text generation, holds promise for optimizing model outputs. However, current evidence on the use of RAG in nursing remains fragmented and has not been systematically synthesized. Objective: This scoping review aims to map the applications of RAG in the nursing field, identify current research characteristics and gaps, and provide a foundational evidence base to inform future research and development of intelligent tools for nursing practice. Methods: This study adopted the Arksey and O’Malley scoping review framework and followed PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) reporting standards. We systematically searched 9 databases—PubMed, Embase, Web of Science, Cochrane Library, Scopus, CNKI, VIP Database, Wanfang Database, and Sinomed—from inception to September 2026. Backward and forward citation searches were also conducted to identify additional eligible studies. Eligible publications were restricted to those in English or Chinese during study selection. Two independent reviewers completed study screening and data extraction, and descriptive synthesis was performed for all eligible records. Results: A total of 28 studies published between 2024 and 2026 were included. The application of RAG in nursing primarily focused on 4 scenarios: patient education (n=9), nursing informatics (n=8), clinical decision support (n=7), and nursing education (n=4). The base LLMs included the GPT series, LLaMA, ChatGLM, and Doubao, among others, while the knowledge bases primarily integrated literature, clinical guidelines, institutional resources, and clinical data. Most studies relied on dense retrieval, while some incorporated hybrid retrieval, knowledge graphs, and reranking to achieve optimization. Generation and task-output evaluations were reported in 27 studies, whereas retrieval evaluation, safety and reliability evaluations, and implementation or outcome evaluations were less frequently reported. Only 4 studies were conducted in real-world clinical settings. Overall, RAG demonstrated potential to improve the accuracy and reliability of LLM-generated outputs in specific nursing tasks; however, most studies validated only in simulated environments. Conclusions: This review systematically maps and synthesizes the application scenarios, technical characteristics, reporting completeness, evaluation methods, and research outcomes of RAG from the perspective of nursing tasks, further refining the evidence landscape for RAG research in nursing contexts. By summarizing the selection of base LLMs, knowledge bases and retrieval processes, the key elements of technical reports, and the areas for tiered evaluation, this review provides evidence-informed considerations for the design, validation, and reporting of future RAG-based nursing AI systems. The findings may assist clinical nurses, nursing educators, and hospital IT teams in identifying application scenarios, key safety requirements, and evaluation priorities, while also highlighting the necessity of conducting prospective validation within real-world nursing workflows.