Author Information
Alexandre Moudjeb, Karlsruhe University of Applied Sciences, GermanyFahmi Bellalouna, Karlsruhe University of Applied Sciences, Germany
Abstract
AI-based assistants are increasingly used in educational contexts, yet large language models (LLMs) often produce inaccurate or hallucinated information, limiting their reliability for learning. This study investigates how Retrieval-Augmented Generation (RAG) can enhance the factuality and coherence of AI responses without additional model fine-tuning. We implement a multimodal educational assistant using RAG, capable of integrating textual documents, code, and images, and evaluating it on a dataset of 2,000 queries sourced from real student interactions and standardized educational exercises. Performance is measured using two complementary metrics: scientific factuality, assessing the accuracy of generated content against verified sources, and multimodal coherence, evaluating consistency across text, code, and images. A novel scoring method based on Jensen-Shannon divergence and temperature softmax is introduced, enabling automatic and dynamic evaluation of the system’s responses. Validation against a set of human-annotated queries confirms the robustness of the scoring framework. Experimental results show that integrating RAG improves response accuracy by 9% and increases overall performance, demonstrating that RAG is a decisive factor in delivering reliable educational content. The multimodal approach further enhances the assistant’s ability to address diverse query types, from textual explanations to code examples and visual illustrations. These findings suggest that RAG-based AI assistants can provide more trustworthy and pedagogically effective support, potentially improving student engagement and learning outcomes.
Paper Information
Conference: ECE2026Stream: Assessment Theories & Methodologies
Published: Wednesday, September 23, 2026
This paper is part of the ECE2026 Conference Proceedings (View)
Full Paper
View / Download the full paper in a new tab/window
To cite this article:
Moudjeb A., & Bellalouna F. (2026) Building an AI Teaching Assistant With RAG: Implementation and Performance Scoring Framework ISSN: 2188-1162 The European Conference on Education 2026: Official Conference Proceedings (pp. 171-187) https://doi.org/10.22492/issn.2188-1162.2026.14
To link to this article: https://doi.org/10.22492/issn.2188-1162.2026.14
Comments
Powered by WP LinkPress