DentalEdu-AI graph: mapping evidence-grounded and knowledge-structured AI for dental education, assessment, and accuracy.

Post-2022 increase of large language models (LLMs) in dental education has generated a expansive heterogeneous literature spanning MCQ benchmarking, OSCE support, curriculum mapping. Despite significant increase, there is no evidence map showing where knowledge is robust or where methodological investment is needed. The function of retrieval-augmented generation (RAG) and knowledge graphs (KGs) in minimizing hallucinations is unexplored. The DentalEdu-AI Graph was created from 667 records: PubMed (549), Scopus (20), and web sources (98). After screening, 447 studies were chosen. Multi-label extraction identified AI type, dental specialty, assessment format, study design, and four quality flags was identified. A 37-node/49-edge knowledge graph of the dental AI ecosystem was constructed and validated with TF-IDF RAG-lite retrieval across 10 curriculum queries. Publication output risen considerably from 2 in 2022 to 203 in 2025 (+ 10,050%). ChatGPT/GPT-3.5 tops with 77.6%, followed by Gemini/Bard at 42.1% and Claude at 16.3%. Accuracy accounts for 80.5% of studies, but hallucination is only focused in 7.8%, revealing a major critical gap. RAG appears in 2.9%, and KG in 0.4%, despite being key hallucination-suppression strategies. The RAG-lite demo accomplished 80% top-1 accuracy, 90% top-3 accuracy, and Cohen's κ = 0.640. Dentistry LLM education literature is obsessed with precision and lacks infrastructure. Quality-first system design using RAG pipelines and KG limitations must replace accuracy-focused benchmarks. A KG (37 nodes, 49 edges, 21 typed relations) and TF-IDF demo are publically available in DentalEdu-AI Graph, the first open evidence map and infrastructure blueprint.