EduBot: Integrating Federated Learning and RAG for Educational Document Analysis
- 1Department of Computer Engineering, Kocaeli University, Kocaeli 41001, Turkey
Abstract
There is a growing demand for intelligent and secure document management systems in the education sector, driven by the exponential growth of digital content. Nevertheless, classic centralized machine learning methods pose major privacy challenges. In this study, we present EduBot, a decentralized intelligent framework for education that integrates Federated Learning (FL) with Retrieval-Augmented Generation (RAG) to ensure privacy and deliver an advanced user experience for content interaction. EduBot introduces a dual-mechanism architecture that enables users to interact with their private documents in real time via a RAG-based conversational interface while also allowing local fine-tuning of a Low-Rank Adaptation (LoRA) adapter on the same data. The masked adapter updates are aggregated on a central server using the Federated Averaging method, enabling collaborative learning without exposing any private documents. The system was evaluated using 29 textbooks published by the Turkish Ministry of National Education (MoNE) and 1318 validated questions. Federated Learning (FL) produced a single unified model that could be deployed across all three domains without sharing the underlying documents. The RAG system retrieved the correct source page among the top five results for 93.3% of the questions. When combined with RAG, the FL model achieved the highest answer quality among the Qwen2.5-1.5B-based models, outperforming the centralized model (token F1 of 0.067 and 0.053, respectively). Overall, our study validates the feasibility of developing privacy-preserving AI assistants that can support educational ecosystems without centralizing sensitive data.
Related
- Data Amount Heterogeneity Improves Performance in Federated Learning for Neuroimaging
- BCFL-DAI-PUB: Blockchain-integrated federated learning with dynamic adaptive incentives for privacy-utility balanced credit risk modelling
- Midwifery students’ and teachers’ perceptions and needs regarding Extended Reality (XR) in educational training and its potential role in anxiety-related learning contexts
- Letter to the Editor: Integrating Machine Learning Into Stroke Risk Stratification in AF—Authors' Reply
- Integrating Multisource Target Evidence with Machine-Learning Potency Prediction for Alzheimer's Disease Drug Discovery
BibTeX
@article{sahin2026edubot,
title = {{EduBot: Integrating Federated Learning and RAG for Educational Document Analysis}},
author = {Begüm Erva Şahin and Murat Görkem Çoban and Ahmet Sayar and Meltem Kurt Pehlivanoğlu},
journal = {Applied Sciences},
year = {2026},
month = oct,
doi = {10.3390/app162010224},
url = {https://doi.org/10.3390/app162010224},
note = {CC BY},
}
RIS
TY - JOUR
TI - EduBot: Integrating Federated Learning and RAG for Educational Document Analysis
AU - Begüm Erva Şahin
AU - Murat Görkem Çoban
AU - Ahmet Sayar
AU - Meltem Kurt Pehlivanoğlu
PY - 2026
DA - 2026/10/26
JO - Applied Sciences
DO - 10.3390/app162010224
UR - https://doi.org/10.3390/app162010224
AB - There is a growing demand for intelligent and secure document management systems in the education sector, driven by the exponential growth of digital content. Nevertheless, classic centralized machine learning methods pose major privacy challenges. In this study, we present EduBot, a decentralized intelligent framework for education that integrates Federated Learning (FL) with Retrieval-Augmented Generation (RAG) to ensure privacy and deliver an advanced user experience for content interaction. EduBot introduces a dual-mechanism architecture that enables users to interact with their private documents in real time via a RAG-based conversational interface while also allowing local fine-tuning of a Low-Rank Adaptation (LoRA) adapter on the same data. The masked adapter updates are aggregated on a central server using the Federated Averaging method, enabling collaborative learning without exposing any private documents. The system was evaluated using 29 textbooks published by the Turkish Ministry of National Education (MoNE) and 1318 validated questions. Federated Learning (FL) produced a single unified model that could be deployed across all three domains without sharing the underlying documents. The RAG system retrieved the correct source page among the top five results for 93.3% of the questions. When combined with RAG, the FL model achieved the highest answer quality among the Qwen2.5-1.5B-based models, outperforming the centralized model (token F1 of 0.067 and 0.053, respectively). Overall, our study validates the feasibility of developing privacy-preserving AI assistants that can support educational ecosystems without centralizing sensitive data.
SN - 2076-3417
ER -