The development of Large Language Models (LLMs) has led to significant advancements in natural language processing and enabled numerous applications across various industries. However, many LLM-based solutions operate as open systems relying on cloud services, which pose risks to data confidentiality and security. To address these challenges, organizations require closed LLM systems that comply with data protection regulations while maintaining high performance. In this paper, we present a reference architecture for developing closed, LLM-based systems using open-source technologies. The architecture provides a flexible and transparent solution that meets strict data privacy and security requirements. We analyze the key challenges in implementing such systems, including computing resources, data management, scalability, and security risks. Additionally, we introduce an evaluation pipeline that enables a systematic assessment of system performance and compliance.
@article{arxiv.2503.01915,
title = {Datenschutzkonformer LLM-Einsatz: Eine Open-Source-Referenzarchitektur},
author = {Marian Lambert and Thomas Schuster and Nico Döring and Robin Krüger},
journal= {arXiv preprint arXiv:2503.01915},
year = {2025}
}