DESIGN AND DIAGNOSTIC EVALUATION OF A RAG-BASED EDUCATIONAL CHATBOT FOR DISTANCE TRAINING IN DIGITAL TECHNOLOGY AND ARTIFICIAL INTELLIGENCE FOR EDUCATION PROFESSIONALS IN SENEGAL

This study examines the design and diagnostic evaluation of an educational chatbot based on retrieval-augmented generation (RAG) for the distance training of education professionals in Senegal. The protocol compares four embedding models on 54 answerable questions and five small language models on 60 questions under two document-context conditions: gold-standard passages and fixed RAG contexts. BGE-M3 achieves the highest overall descriptive scores for dense retrieval (Hit@3 = 0.500; MRR@10 = 0.448), although performance varies by question type. For all five generators, replacing the RAG context with the gold context improves automatically assessed accuracy and coverage of the expected information. The findings highlight the interdependence of corpus organization, retrieval, and generation, without establishing the pedagogical superiority of any model.