RAGAETON: Retrieval Augmented Generative Ai Education TOol with Networks

RAGAETON

Nicolas Bioul, UCLouvain Corentin Warzée, Uclouvain Thomas Baldassarre, UMons

Training / education / pedagogy

A university deploys an AI learning companion that transforms course materials into validated concept maps and provides traceable answers. Students use it for studying and self-testing, while teachers oversee content and adaptation. AI intervenes during learning tasks to personalize difficulty and explain reasoning, balancing efficiency with transparency and data privacy concerns.

This project matters because it directly addresses a growing structural tension in education: the rapid adoption of AI tools is outpacing our ability to ensure they are trustworthy, pedagogically sound, and socially acceptable. Many current AI systems optimize for performance and convenience, but they often neglect core educational requirements such as explainability, epistemic rigor, and learner agency. RAGAETON intervenes precisely at this gap by reframing AI not as an answer machine but as a mediator of structured knowledge. In today’s context, trust is a central issue. Learners and educators are increasingly skeptical of systems that produce fluent outputs without clear justification. The emphasis on traceability and source inspection responds to a broader demand for epistemic transparency, especially in an era shaped by misinformation and generative AI hallucinations. By allowing users to inspect reasoning pathways and underlying sources, the project aligns AI with academic norms of citation, verification, and critique, which are foundational to education but often absent in mainstream tools. Equally important is the question of control. Many adaptive learning systems operate as black boxes, adjusting difficulty or content in ways that can feel arbitrary or disempowering. In contrast, RAGAETON restores agency by making adaptation tunable by both learners and teachers. This is particularly relevant today, as personalization technologies risk drifting into forms of soft determinism, where users are passively shaped by algorithmic decisions rather than actively participating in their learning trajectory. The project also responds to increasing regulatory and ethical pressures, especially in regions governed by frameworks like GDPR and emerging AI regulations. Privacy is no longer a secondary feature but a design constraint, and systems that fail to integrate it at the architectural level face both legal and societal resistance. By embedding privacy controls, data minimization, and user rights (such as deletion and export), RAGAETON anticipates the direction of compliant AI systems rather than retrofitting safeguards after deployment. Another reason this project is timely lies in its approach to knowledge representation. Education is not simply about accessing information but about understanding how concepts relate to one another. Most digital tools still treat knowledge as flat text or isolated resources. By transforming documents into concept graphs and linking them to semantic retrieval systems, RAGAETON aligns with cognitive science insights about how humans build mental models. This makes it particularly relevant in a context where information overload is a defining challenge: structuring knowledge becomes more valuable than merely retrieving it. Finally, the project matters because it redefines the role of AI in education at a systemic level. Rather than remaining a peripheral assistant, it becomes part of a pedagogical infrastructure that includes validation workflows, governance mechanisms, and bias mitigation. This shift is critical today, as educational institutions are not just experimenting with AI tools but beginning to integrate them into curricula, assessment, and institutional processes. Systems that fail to account for these broader implications risk undermining both educational quality and institutional trust. In sum, RAGAETON is relevant because it responds to the central challenges of contemporary AI in education: trust, control, privacy, structure, and governance. It does not simply improve existing tools; it proposes a different paradigm in which AI supports learning by making knowledge explicit, navigable, and accountable.

Reforming and personalizing the way students learn and subjects are taught

RAGAETON is designed from the outset around a clear premise: effective AI-mediated learning must align with the explicit expectations of both learners and educators. Interviews conducted with these groups highlight several non-negotiable requirements, including transparent reasoning with accessible sourcing, strict privacy guarantees compliant with GDPR, and adaptive learning systems that remain under user control. At the same time, users express a preference for optional, rather than imposed, social interaction, alongside concerns about excessive gamification, bias reinforcement, and environmental impact. What they value most is not automation alone, but structured representations of knowledge, explainable recommendations, and the ability to explore multiple perspectives through alternative learning agents. To address these needs, RAGAETON is conceptualized as more than a simple tool. It evolves across three levels of functionality. Initially, it operates as a tool providing discrete assistance, but it progressively becomes a learning companion capable of continuous interaction and adaptation. Ultimately, it functions as a full pedagogical infrastructure, embedding governance mechanisms, validation workflows, and ethical constraints directly into the system. This progression reflects a shift from isolated functionality toward an integrated, accountable learning environment. The system is built in response to well-documented shortcomings of traditional learning management systems. These platforms typically provide access to content but fail to explain how knowledge is structured, why specific answers are given, or how difficulty evolves over time. They often lack transparency in sourcing, offer limited control over adaptive mechanisms, impose social dynamics that may not suit all learners, and rely on coarse retrieval methods that obscure conceptual relationships. Additionally, users frequently perceive their handling of sensitive data as inadequate, while their tendency to present simplified or flattened knowledge raises concerns about bias. RAGAETON addresses these limitations through three core design principles. First, traceability ensures that every generated answer is explicitly linked to validated sources, reinforcing trust and accountability. Second, structured conceptualization transforms traditional documents into interactive concept graphs, allowing learners to navigate knowledge as an interconnected system rather than as isolated fragments. Third, personalized adaptability enables both learners and teachers to fine-tune parameters such as difficulty, tone, and pedagogical focus, ensuring that the system remains responsive without being prescriptive. Technically, these principles are implemented through a hybrid architecture combining retrieval-augmented generation, teacher-validated knowledge graphs, multimodal embeddings, adaptive evaluation heuristics, and user-centered privacy controls. The system begins by ingesting raw pedagogical materials, which are normalized and enriched with metadata before being transformed into a structural graph representing documents, chapters, and semantically coherent text chunks. In parallel, a multimodal embedding space is constructed, encoding both textual and visual content into a shared vector representation that supports precise semantic retrieval. On top of this structural and vector foundation, the system extracts educational concepts using a language model, identifying key notions, categorizing them, and establishing relationships between them. A two-stage process refines these relationships to ensure coherence and eliminate inconsistencies, resulting in a concept graph that reflects the logical structure of the material. Crucially, this graph is not deployed directly. Instead, it undergoes a teacher validation phase, during which educators can modify, correct, or enrich the representation. This step ensures academic rigor, mitigates bias, and aligns the system’s outputs with pedagogical intent. Once validated, the system becomes accessible to learners through two main modules. The learning module includes a chatbot capable of generating grounded explanations, a flashcard system for focused practice, and an interactive concept map that reveals the global structure of the subject matter. The testing module complements this by offering open-ended questions evaluated according to teacher-defined heuristics, as well as graph-completion tasks that require learners to reconstruct conceptual relationships. Together, these tools emphasize deep understanding over rote memorization. In its entirety, RAGAETON transforms unstructured educational documents into navigable, interactive, and trustworthy knowledge environments. By integrating explicit graph structures, multimodal semantic retrieval, and teacher oversight, it shifts the role of AI in education from answer generation to knowledge mediation. The system thus embodies a transition from content delivery to structured understanding, from opaque automation to explainable reasoning, and from isolated tools to a governed pedagogical infrastructure.