EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

EduRAG: A Multimodal Explainable RAG System With Deterministic Token Grounding

S. Amal, A. R. Niranjan, M. G. Thushara

IEEE Access · 2026 · 14 · 90849-90859

DOI: 10.1109/access.2026.3704123 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von S. Amal und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

LLMs (Large Language Models) offer quick access to information, but they tend to be unreliable when used for academic purposes. Because of this, they often generate factually incorrect information (known as factual hallucinations) and do not often provide reliable citations. In this work, we introduce EduRAG, a multimodal, explainable retrieval-augmented generation model, specifically engineered to build trust in the field of education. It includes a hybrid information retrieval system incorporating the advantages of the BM25 lexical match, combined with FAISS vector space embeddings, and fine-tuned via Hypothetical Document Embeddings (HyDE) query expansion and cross-encoder re-ranking on MS-MARCO as a superior alternative to other methods. To address the constraints arising from the high computational cost associated with conventional attribution methods, such as SHAP values, we introduce a multi-level attribution framework that enables the transparent inspection of evidence on the chunk level with cross-encoder evidence scores, deterministic token-level grounding heatmaps for hallucination detection, and source disagreement for contradiction spotting between different retrieved sources. Controlled experimental results conducted with a customized dataset consisting of 97 educational queries demonstrate that EduRAG achieves a Mean Reciprocal Rank (MRR) score of 0.7784 and Recall@5 score of 0.8763. A targeted visual evaluation on 15 diagram-specific queries demonstrates visual MRR of 0.967 and Recall@5 of 100%, validating VLM-enabled multimodal retrieval. The answers generated by the model yield a BERTScore F1 of 0.8952, faithfulness score of 0.8924, a 94% citation validity rate with verifiable provenance, and 85.5% token grounding accuracy. The impact of each component in the proposed system is validated via an ablation study involving three retrieval-based models.

Zitieren

APA

Amal, S., Niranjan, A. R., & Thushara, M. G. (2026). EduRAG: A Multimodal Explainable RAG System With Deterministic Token Grounding. IEEE Access, 14, 90849-90859. https://doi.org/10.1109/access.2026.3704123

Vancouver

Amal S, Niranjan AR, Thushara MG. EduRAG: A Multimodal Explainable RAG System With Deterministic Token Grounding. IEEE Access. 2026;14:90849-90859. doi:10.1109/access.2026.3704123

BibTeX
@article{Amal2026,
  author = {Amal, S. and Niranjan, A. R. and Thushara, M. G.},
  title = {EduRAG: A Multimodal Explainable RAG System With Deterministic Token Grounding},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {90849--90859},
  doi = {10.1109/access.2026.3704123}
}
Kostenlos starten →