Moodle's new RAG-based AI tutor achieves 0.97 faithfulness score
This open-source plugin cuts hallucinations with teacher-grounded Socratic tutoring.
A team of Polish researchers from Warsaw University of Technology has published details of their AI Teaching & Learning Assistant, a modular plugin for the open-source LMS Moodle. The system uses Retrieval-Augmented Generation (RAG) to ground Large Language Model responses exclusively in teacher-provided course materials, virtually eliminating hallucinations—a frequent criticism of AI in education.
The plugin offers a dual-centric design: students engage with an interactive, Socratic-questioning tutor that promotes deep conceptual understanding rather than surface-level answers. Educators, meanwhile, get a human-in-the-loop workspace to review, edit, and approve AI-generated content before it reaches learners. In evaluations using the Ragas (LLM-as-a-Judge) framework, the system scored 0.97 out of 1.0 on faithfulness, and a preliminary user study yielded a 4.00/5.00 recommendation rate. The paper was accepted as a demo at IJCAI 2026.
- Plugin uses Retrieval-Augmented Generation (RAG) to ground LLM responses in teacher-uploaded materials, achieving a 0.97 faithfulness score.
- Students interact via Socratic-questioning tutor; educators have a human-in-the-loop content approval workflow.
- Accepted as a demo paper at IJCAI 2026; 4.00/5.00 recommendation rate from preliminary user study.
Why It Matters
Brings reliable, hallucination-free AI tutoring to Moodle—the world's most widely used LMS—empowering educators to scale personalized learning.