AI-based assistants are increasingly playing a central role in educational environments, yet pre-trained large language models (LLMs) remain limited by issues of factuality and hallucinations, which are particularly critical in pedagogical contexts. To address these challenges, this paper investigates the contribution of Retrieval-Augmented Generation (RAG) systems, which combine text generation with information retrieval from external knowledge sources, without requiring additional model fine-tuning. We propose a systematic implementation of an educational AI assistant based on FastAPI, designed following object-oriented programming principles. The system integrates a multimodal RAG pipeline leveraging textual documents, code, and images. Its performance is evaluated on a dataset of 2,000 user queries, comparing state-of-the-art multimodal generation models (Ministral-3:14b), with and without the RAG mechanism. A simplified scoring method is introduced, based on scientific factuality and multimodal coherence, enabling analysis of the combined impact of dynamically provided documents and those indexed in the RAG knowledge base. Experimental results indicate that integrating RAG improves response accuracy by 9% and significantly boosts the overall performance score, demonstrating that RAG is a decisive lever for deploying reliable and pedagogically effective AI assistants in education. Our innovative scoring method, based on Jensen-Shannon divergences and temperature softmax, which automatically updates and allows evaluation of the performance of an object-oriented AI assistant, taking into account the multimodality of sources; in our case, it achieved a score of 7.2/10.

Authors:
Alexandre Moudjeb, Hochschule Karlsruhe, Germany
Fahmi Bellalouna, Hochschule Karlsruhe, Germany

About the Presenter(s)
Alexandre Moudjeb, a laboratory assistant and Franco-German student, passionate about new technologies, including AI and its applications.

See this presentation on the full schedule – On Demand Schedule

Bookmark