Rag in practice : Multi domain applications for academic and municipal services

Abstract

Retrieval-Augmented Generation (RAG) systems are increasingly applied in various domains to improve access to information and improve user interaction with complex knowledge sources. However, most prior research remains limited to benchmark evaluations or theoretical designs, with little practical evidence of how RAG performs in real-world, domain-specific contexts requiring transparency and accessibility. To address this gap, we studied the practical implementation of RAG in three domains: education, research, and municipal governance through the development of NoteScholar, ThesisBuddy, and CityOfficeRAG. Each system was designed to meet different user needs. We followed the Design Science Research (DSR) framework for the systems. NoteScholar supports students and independent learners by combining a document-based RAG chatbot with a configurable podcast generator. ThesisBuddy provides teachers and students with a user-friendly File management system( through the admin panel) based on a RAG environment for managing and querying academic information. CityOfficeRAG adapts the framework to a civic environment, which allows administrative staff to query and interpret municipal regulations. All three systems integrate document chunking, embedding-based retrieval, re-ranking, LLM-driven response generation, and explicit source attribution to ensure reliable outputs.

We evaluated the systems using a multi-method approach. A user survey with 10 participants (5 researchers and 5 students) assessed NoteScholar’s usability and learning impact, while ThesisBuddy and CityOfficeRAG were reviewed by field-specific experts for retrieval accuracy and transparency. Results showed that 90\% of users found document upload and querying intuitive, 80\% rated retrieval accuracy as “good” or better, and 9 out of 10 users reported improved comprehension through podcast-based engagement. Expert feedback further confirmed the framework’s adaptability and clarity across research and governance contexts. This thesis makes three key contributions:(1) the design and deployment of a unified, domain-adaptable RAG framework; (2) Practical insights into usability, transparency, and trust in applied RAG systems; and (3) methodological guidance for low-resource, real-world RAG deployments.

Our findings demonstrate that RAG systems, when designed for modularity and transparency, can provide accessible and trustworthy information services. However, this work also faces several limitations, including a small dataset and a limited participant sample size, which may affect scalability and cost-efficiency. Future work will address these challenges through larger-scale evaluations, integration of open-source or multilingual models, and studies on long-term system adaptability and policy drift handling.