RAGAETON: Retrieval Augmented Generative Al Education TOol with Networks

Team name

RAGAETON

Use of AI tools

• ChatGPT

For what purposes?

The AI tools were used for:

• Literature exploration and synthesis

• Structuring ideas and refining arguments

• Brainstorming architectural and pedagogical concepts

• Generating alternative perspectives and ethical scenarios

• Supporting drafting and reformulation of sections

• Simulating “synthetic user” personas during the ethical evaluation phase

• Improving clarity, coherence, and academic readability

How did these tools support the work?

AI tools acted primarily as cognitive support systems rather than autonomous authors. They helped accelerate ideation, compare conceptual alternatives, identify gaps in reasoning, and improve the structure of the proposal. Generative AI was also experimentally integrated into the research methodology itself through the use of AI-generated personas during the ethical evaluation phase. However, all conceptual decisions, system architecture, ethical positioning, synthesis, and final validation remained human-driven and were critically reviewed by the team.

External feedback & contributions

Alexis Cambriel — Student / Participant — Feedback on teacher involvement and moderation

Noël Junior Yando Fotso — Student / Research contributor — Critical discussion on scalability, teacher workload, and institutional feasibility

Haziel Álvarez — Participant / Reviewer — Feedback on governance and risks of superficial validation

Youri (ÒCTAVIA Team) — Participant / Reviewer — Discussion on operational feasibility and validation bottlenecks

Bruno De Lièvre — Professor, Faculty of Psychology and Educational Sciences, University of Mons — Expert feedback on realism, accessibility, implementation strategy, and pedagogical framing

Joshua Kolawole — Participant / Reviewer — General feedback on educational value, trust, and architectural coherence

Nipun Ranchhod Navadia — Participant / Reviewer — Critical discussion on sustainability, scalability, and institutional maintenance

Initial contribution

RAGAETON: Retrieval Augmented Generative Ai Education TOol with Networks

Final contribution

The final contribution builds upon our original proposal for RAGAETON (Retrieval Augmented Generative AI Education Tool with Networks), an AI-mediated educational system designed to transform pedagogical materials into validated, navigable, and explainable knowledge structures.

Our work evolved significantly during Phase 2. Initially, the project focused primarily on the technical architecture: retrieval-augmented generation, multimodal embeddings, concept extraction, teacher validation, and adaptive learning mechanisms. Through discussions, feedback, and expert critiques, the project progressively shifted from being conceived as a “smart learning assistant” toward a broader pedagogical infrastructure grounded in governance, transparency, and human-centered AI design.

The strongest influence on this evolution came from recurring concerns regarding scalability, teacher workload, and institutional realism. Several contributors questioned whether requiring teachers to validate concept graphs and pedagogical structures would create an unsustainable burden. These discussions pushed us to rethink validation not as continuous manual correction, but as a lightweight supervisory process integrated into existing pedagogical workflows. As a result, the proposal evolved toward a model where AI generates provisional structures that educators review, adjust, and reuse collaboratively over time rather than rebuilding from scratch.

Another major influence was the feedback emphasizing the importance of making the student experience concrete and relatable. This encouraged us to reposition the project around a realistic educational scenario: students using AI to revise complex subjects while remaining uncertain about the reliability, origin, and pedagogical quality of generated answers. This shift strengthened the clarity of the proposal by grounding abstract technical mechanisms in a visible educational problem centered on trust and epistemic transparency.

The ethical and regulatory dimensions of the project also became substantially more developed during this phase. Discussions around GDPR compliance, data governance, learner profiling, bias reinforcement, and environmental sustainability motivated the integration of a formal human-centered ethical framework inspired by the AI Dilemma Cards methodology. Rather than treating ethics as a secondary compliance layer, the final proposal embeds ethical reflection directly into system design, evaluation, and governance processes.

Another important evolution concerned the role of AI itself. Early versions of the project risked presenting AI primarily as an optimization mechanism. Through feedback and reflection, the proposal progressively reframed AI as a mediator of structured knowledge rather than an autonomous answer machine. This conceptual shift became central to the identity of RAGAETON. The system’s purpose is no longer simply to generate responses efficiently, but to help learners understand how concepts relate, why answers emerge, and where knowledge originates.

The process also reinforced the importance of transparency and explainability at every level of the architecture. The final proposal therefore emphasizes:

• traceable retrieval pipelines,

• teacher-validated concept graphs,

• explicit source inspection,

• adaptive learning controls that remain user-governed,

• and transparency layers allowing learners to inspect reasoning pathways.

Technically, the final system combines:

• retrieval-augmented generation (RAG),

• multimodal semantic embeddings,

• concept extraction pipelines,

• structural and conceptual knowledge graphs,

• adaptive learning modules,

• and teacher-supervised evaluation heuristics.

The project ultimately evolved from a technically ambitious educational AI tool into a broader reflection on how AI can be responsibly integrated into higher education without sacrificing trust, learner agency, pedagogical rigor, or institutional accountability.

One of the major outcomes of this process was recognizing that educational AI systems cannot be evaluated solely according to performance metrics such as speed, convenience, or answer accuracy. Instead, they must also be assessed according to their cognitive, ethical, social, and regulatory implications. This realization strongly shaped the final version of the proposal and strengthened its interdisciplinary foundation.

The final paper therefore proposes not only a technical system, but also a human-centered framework for designing and evaluating educational generative AI systems under real pedagogical and institutional constraints.

Reflection on the process

The collaborative and iterative nature of Phase 2 was essential in strengthening the project. Feedback from participants, reviewers, and experts repeatedly challenged our assumptions regarding feasibility, governance, scalability, and pedagogical realism. Rather than weakening the proposal, these critiques helped clarify its priorities and exposed the tensions that educational AI systems must address in practice.

Several discussions highlighted that the true challenge was not simply building a technically capable AI assistant, but designing a system that institutions, teachers, and learners could realistically trust and adopt. This shifted our attention from purely functional innovation toward governance, explainability, and institutional integration.

The process also demonstrated the importance of interdisciplinary thinking. Educational AI cannot be approached exclusively as a technical problem; it simultaneously involves pedagogy, ethics, cognitive science, regulation, accessibility, and human-computer interaction. Integrating these dimensions made the proposal significantly more coherent and realistic.

Finally, the iterative exchanges revealed the value of criticism itself. Concerns about teacher workload, validation bottlenecks, sustainability, and learner dependence forced the project to become more precise, more transparent about limitations, and more grounded in operational reality. The final proposal is therefore substantially stronger, more balanced, and more mature than the initial contribution because it emerged through continuous confrontation between technical ambition and educational responsibility. Link to paper