JMIR Publications

ABSTRACT

Background:

The rapid development of large language models (LLMs) has accelerated the adoption of generative artificial intelligence in digital health. Despite their potential, LLMs face critical limitations, including hallucinations, insufficient domain-specific knowledge, and limited integration of up-to-date clinical evidence, which constrain their application in nursing practice. Retrieval-augmented generation (RAG), a technique that combines information retrieval with text generation, has emerged as a promising solution to improve the accuracy, transparency, and evidence alignment of AI systems. However, current evidence on the use of RAG in nursing remains fragmented and has not been systematically synthesized.

Objective:

To systematically map the applications of retrieval-augmented generation (RAG) in the nursing field, identify current research hotspots and gaps, and provide evidence to inform the development of intelligent tools and clinical practice in nursing.

Methods:

This study was conducted based on the Arksey and O’Malley scoping review framework and reported in accordance with the PRISMA-ScR guidelines. A comprehensive search was performed in both English and Chinese databases, including PubMed, Embase, Web of Science, and CNKI, from database inception to March 2026. Two reviewers independently conducted study selection and data extraction, and a descriptive analysis of the included studies was performed.

Results:

A total of 26 studies published between 2024 and 2026 were included, covering multiple countries such as China, the United States, and Australia. The applications of RAG in nursing were primarily concentrated in patient education and self-management, followed by clinical decision support, nursing information processing, and nursing education. In terms of technical architecture, the underlying models included the GPT series, LLaMA, ChatGLM, and Doubao. Knowledge bases were mainly derived from literature and clinical guidelines. Dense retrieval was the dominant retrieval strategy, while some studies incorporated hybrid retrieval, knowledge graphs, or reranking mechanisms for optimization. Evaluation methods mainly focused on technical performance, content quality, and user experience. Overall, application outcomes were positive, with RAG improving the accuracy and stability of model outputs and narrowing performance differences among models.

Conclusions:

RAG demonstrates promising potential in the nursing field. However, current research remains in an exploratory stage, with limitations including uneven application scenarios, insufficient knowledge base quality and standardization, lack of unified evaluation frameworks, and limited clinical evidence. Future research should expand application scenarios, optimize technical architectures, standardize evaluation systems, and strengthen validation in real-world clinical settings to promote the standardized and high-quality development of RAG in nursing practice.