Application of Retrieval-Augmented Generation Large Language Models in Stomatology: A Scoping Review

Abstract

Large language models (LLMs) have been gradually applied to tasks such as specialized question answering, educational assistance, clinical decision support, and knowledge integration in stomatology. However, general-purpose LLMs still suffer from critical limitations, including outdated knowledge, hallucinatory generation, untraceable information sources, and inadequate clinical adaptability. Retrieval-augmented generation (RAG) offers a core solution for improving the accuracy, interpretability, and evidential traceability of model outputs by integrating external authoritative knowledge sources. Despite the rapid growth of relevant research, a systematic review on the application status of RAG-enhanced LLMs in the field of stomatology remains lacking. This is a scoping review reported in accordance with the PRISMA-ScR guidelines. A systematic search was performed in the PubMed, Web of Science, Embase, Scopus, and IEEE Xplore databases from January 2023 to August 2026. Original studies that developed, applied, or evaluated RAG-enhanced large language models in the field of stomatology were included. Literature screening and data extraction were independently conducted by 2 researchers. Extracted items included research scenarios, subspecialties, knowledge sources, RAG technical characteristics, evaluation design, and main findings. A narrative synthesis approach was used to integrate the results. A total of 20 studies were included, most of published between 2025 and 2026, covering many subspecialties of stomatology. Application tasks included specialized question-and-answer and evidence-based support, clinical decision making and diagnostic assistance, treatment or prescription recommendation, as well as education and training support. Existing studies consistently demonstrated that RAG-enhanced LLMs significantly improved knowledge relevance, evidential traceability, and task adaptability in oral specialty settings, and outperformed unenhanced generic LLMs in multiple comparative evaluations. However, current evidence remains dominated by prototype development, simulated scenario assessments, and small sample studies, with substantial heterogeneity across studies in knowledge base construction, retrieval methods, evaluation metrics, and the extent of clinical validation. RAG-enhanced LLMs show preliminary potential in stomatology, particularly in task scenarios requiring specialized knowledge grounding, evidence-based support, and traceable information sources. However, the field remains at an early stage of development, and current evidence is largely based on prototype systems, benchmark datasets, simulated scenarios, and limited validation settings. Further studies are needed to strengthen real-world validation, multimodal knowledge integration, standardized evaluation frameworks, and safety governance mechanisms before this technology can be responsibly translated into stomatological practice.