A METHOD FOR SELECTING THE ARCHITECTURE OF LLM, RAG, AND AI AGENT DEPLOYMENT IN CORPORATE INFORMATION SYSTEMS AND DECISION SUPPORT SYSTEMS
DOI:
https://doi.org/10.14529/ctcr260301Keywords:
large language models, LLM, Retrieval-Augmented Generation, RAG, AI agents, corporate information systems, decision support systemsAbstract
Purpose of the study. The purpose of the study is to develop a method for selecting the deployment architecture of LLM subsystems, Retrieval-Augmented Generation (RAG) solutions, and AI agents in corporate information systems and decision support systems. The study is relevant because in the enterprise environment the choice of an intelligent subsystem depends not only on generation quality, but also on explainability, controllability, integration depth, work with corporate data, acceptable autonomy, and the cost of error. Materials and methods. The study uses system and comparative analysis, architectural modeling, classification, expert decomposition of criteria, and elements of multicriteria decision-making. Five architectural classes are considered: an isolated LLM subsystem, a RAG-based subsystem, an LLM contour with tool access, an agent-based intelligent subsystem, and a hybrid controlled architecture. Comparison criteria include task type, dependence on corporate context, need for actions in external systems, requirements for accuracy, explainability and auditability, data sensitivity, enterprise environment maturity, and acceptable autonomy. Results. A semi-formal architecture selection method is proposed. It combines filtering and ranking criteria, weight adjustment, and the possibility of determining an intermediate architecture when data and integrations are insufficiently mature. The method is implemented as a reproducible procedure including task profile formation, filtering by critical constraints, evaluation of admissible architectures, and refinement of the choice according to enterprise maturity. Its applicability is demonstrated using scenarios of a trade-and-production enterprise: access to internal regulations, managerial reporting, recommendation generation for a standard customer request, and multi-step handling of an internal incident. Conclusion. There is no universal deployment architecture for an intelligent subsystem in corporate information systems and decision support systems. A rational choice is determined by technological, organizational, infrastructural, and managerial factors. The proposed method can be used for preliminary design and phased development of intelligent subsystems in organizations focused on automating managerial and information-analytical processes.References
1. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks / P. Lewis, E. Perez, A. Piktus et al. // ArXiv preprint: arXiv:2005.11401. 2020.
2. A Survey on Large Language Model Based Autonomous Agents / L. Wang, C. Ma, X. Feng et al. // ArXiv preprint: arXiv:2308.11432. 2023.
3. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions / L. Huang, W. Yu, W. Ma et al. // ArXiv preprint: arXiv:2311.05232. 2023.
4. Добренко Н.В., Добренко Д.А., Улизько М.В. Интеллектуальная поддержка принятия управленческих решений в MES-системах с использованием больших языковых моделей // Экономика. Право. Инновации. 2024. № 3. С. 47–59. DOI: 10.17586/2713-1874-2024-3-47-59
5. Бурый А.С., Цаплина О.С. Агентные модели принятия решений на основе генеративного искусственного интеллекта // Правовая информатика. 2025. № 1. С. 101–109. DOI: 10.24412/1994-1404-2025-1-101-109
6. Синякова Е.А. Стратегии внедрения ИИ в редакциях местных газет: типология и организационный контекст // Вестник Московского университета. Серия 10. Журналистика. 2025. № 6. С. 80–113. DOI: 10.55959/msu.vestnik.journ.6.2025.80113
7. Использование технологий искусственного интеллекта в российских медиа и журналис¬тике / С.Г. Давыдов, А.В. Замков, М.А. Крашенинникова, М.М. Лукина // Вестник Московского университета. Серия 10. Журналистика. 2023. № 5. С. 3–21. DOI: 10.30547/vestnik.journ.5.2023.321
8. Зуйкина К.Л., Разумова Д.В. Технологии искусственного интеллекта в работе российских СМИ: инструменты, проблемы и угрозы // Вестник Московского университета. Серия 10. Журналистика. 2025. № 5. С. 79–103. DOI: 10.55959/msu.vestnik.journ.5.2025.79103
9. Velasquez M., Hester P.T. An Analysis of Multi-Criteria Decision Making Methods // International Journal of Operations Research. 2013. Vol. 10, no. 2. P. 56–66.
10. Ishizaka A., Nemery P. Multi-Criteria Decision Analysis: Methods and Software. Chichester: John Wiley & Sons, 2013. DOI: 10.1002/9781118644898
11. Saaty T.L. Decision Making with the Analytic Hierarchy Process // International Journal of Services Sciences. 2008. Vol. 1, no. 1. P. 83–98. DOI: 10.1504/IJSSCI.2008.017590
12. Tyree J., Akerman A. Architecture Decisions: Demystifying Architecture // IEEE Software. 2005. Vol. 22, no. 2. P. 19–27. DOI: 10.1109/MS.2005.27
13. Jansen A., Bosch J. Software Architecture as a Set of Architectural Design Decisions // Proceedings of the 5th Working IEEE/IFIP Conference on Software Architecture. 2005. P. 109–120. DOI: 10.1109/WICSA.2005.61
14. VladSpace. Как я построил Graph RAG систему с точностью 96.7% за 5 дней: от научных статей до production-ready пайплайна // Хабр. 24.02.2026. URL: https://habr.com/ru/articles/1003064/ (дата обращения: 23.04.2026).
15. Toolformer: Language Models Can Teach Themselves to Use Tools / T. Schick, J. Dwivedi-Yu, R. Dessì et al. // ArXiv preprint: arXiv:2302.04761. 2023.
16. ReAct: Synergizing Reasoning and Acting in Language Models / S. Yao, J. Zhao, D. Yu et al. // ArXiv preprint: arXiv:2210.03629. 2022.
17. Holistic Evaluation of Language Models / P. Liang, R. Bommasani, T. Lee et al. // ArXiv preprint: arXiv:2211.09110. 2022.
