Download PDFOpen PDF in browserAndarax: Channeling the Flow of Knowledge Through a Secure Distributed LLM Architecture12 pages•Published: June 18, 2026AbstractWe present Andarax, a Retrieval-Augmented Generation (RAG) system designed and deployed at the Universitat Autònoma de Barcelona (UAB) to serve as the foundation for the implementation of AI initiatives within the university. This paper describes the system architecture, the RAG pipeline, the infrastructure challenges encountered during development, and the practical lessons learned from building an LLM-based system within the constraints of a university environment, offering guidance for other institutions seeking to develop similar systems.Andarax is built on a distributed architecture of seven virtual machines and supports dual LLM inference (Qwen2.5-32B and Mistral-7B, both AWQ-quantized and deployed on NVIDIA L40S GPUs), multilingual output in Catalan, Spanish, and English, and real-time streaming responses. As an initial use case, currently deployed for internal use and undergoing evaluation, the system has been configured as an AI teaching assistant for university courses. It provides students and teaching staff with accurate, citation-backed answers derived from course materials (lecture slides, PDFs, and notes), while operating entirely on university-owned infrastructure to ensure data sovereignty and GDPR compliance. Keyphrases: data sovereignty, educational technology, generative ai, large language models, multilingual nlp, on premise ai, retrieval augmented generation (rag) In: Laurence Desnos, Carmen Diaz, Janina Mincer-Daszkiewicz, Lazaros Merakos, Raimund Vogl, Stuart McLellan and Ulrike Lucke (editors). Proceedings of EUNIS 2026 Annual Congress, vol 109, pages 184-195.
|
