Dialogical Learning Support in RAG-Based E-Learning

This paper presents a web-based platform designed to support dialogical learning through a Retrieval-Augmented Generation (RAG) architecture. The system integrates retrieval grounding, context-aware dialogue management, and a modular, model-agnostic design to enable controlled and pedagogically aligned learning supported by Artificial Intelligence (AI) and based on instructor-verified educational materials. The proposed approach supports multilingual interaction, including operation in lower-resource languages such as Bulgarian, and models learning as a continuous dialogue rather than a sequence of isolated queries. To ensure reliable knowledge access, the system employs a hybrid retrieval strategy combining semantic embeddings with lexical matching within a two-stage indexing and retrieval framework. The approach is supported by an empirical evaluation based on a manually constructed question set with human-validated relevance assessment. The results demonstrate that the selected configuration achieves 90% retrieval accuracy at TOP-5 and up to 91.4% at TOP-6, providing a reliable contextual basis for response generation. A complementary manual evaluation of generated responses further indicated strong practical usefulness and generally grounded answer quality. The platform is further designed in alignment with European regulatory principles, emphasizing transparency, traceability, and controlled use of AI in educational environments. Overall, the study demonstrates that integrating retrieval precision with pedagogical structure enables the development of AI systems that support structured and contextually grounded learning processes.