Survey maps how retrieval-augmented generation is reshaping biomedical AI

Large language models in biomedicine face a fundamental conflict between static parameter knowledge and the dynamic nature of clinical evidence, and a new survey argues that retrieval-augmented generation (RAG) has become the field’s leading answer to that problem. The survey, published open access in Artificial Intelligence Review, synthesizes the biomedical RAG landscape from 2020 to 2025 through a structured literature mapping of 178 screened records, from which the authors identified 94 representative studies.

To organize this rapidly growing body of work, the researchers classified the studies into three architectural categories. Naive RAG accounts for 38 studies, or 40.4 percent of the sample. Advanced RAG is the largest group, with 46 studies representing 48.9 percent. Modular RAG remains comparatively rare, with 10 studies, or 10.7 percent.

RAG works by grounding a model’s generation in external data rather than relying solely on what the model learned during training. This grounding is particularly valuable in medicine, where clinical evidence changes continuously and static model parameters can quickly become outdated. However, the survey notes that the approach introduces new complexities in latency and architecture, meaning that retrieving and integrating external information comes at a real computational cost.

One of the survey’s central contributions is the formalization of what the authors call the biomedical RAG trilemma: the inherent trade-offs between reasoning depth, inference latency, and data privacy that constrain current clinical deployment.

The trilemma is visible in the survey’s analysis of recent agentic workflows. These systems enhance diagnostic reasoning, but the survey finds they risk prohibitive latency, a serious concern in clinical settings where timely responses matter. The authors also observe that privacy constraints dictate the choice between powerful cloud-based models and local deployment, forcing institutions to weigh model capability against the sensitivity of patient data.

Looking ahead, the survey identifies an alignment gap in multimodal RAG, where systems that combine text with other data modalities still face challenges in aligning retrieved information with generated output. The authors propose future directions for self-correcting, verifiable clinical agents as a path toward more trustworthy deployment.

Subject of Research: A survey of retrieval-augmented generation technologies, datasets, and clinical applications in biomedicine

Article Title: Retrieval-augmented generation in biomedicine: a survey of technologies, datasets, and clinical applications

Article References: He, J., Zhang, B., Rouhizadeh, H., Chen, Y., Yang, R., Lu, J., Chen, X., Liu, N., & Teodoro, D. (2026). Retrieval-augmented generation in biomedicine: a survey of technologies, datasets, and clinical applications. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11743-y

Image Credits: AI Generated

DOI: 10.1007/s10462-026-11743-y

Keywords: retrieval-augmented generation, large language models, biomedicine, clinical applications, artificial intelligence, survey, data privacy, latency, diagnostic reasoning, multimodal RAG, Retrieval-augmented, generation