Аnalyzing system development for technical and organizational documents compliance with regulatory requirements of the nuclear industry based LLM technology

Abstract

This paper discusses the design, development and testing of a prototype of an application system using artificial intelligence technologies for automated analysis of nuclear industry documentation, including verification of compliance with regulatory documents and regulatory requirements. The system uses large language model (LLM) and search-augmented generation (RAG) technologies. Options for using this technology stack for the use of specialized knowledge domains (an array of regulatory documents and technical instructions of the nuclear industry) for high-quality operational examination of working documents using artificial intelligence are considered. Improving the quality and speed of qualified expertise for complex high-tech industries can often be critical for the implementation of serious projects.

Due to the availability of significant sensitive information of a technical and economic nature in the industry, it is important to organize work without external access to the analyzed design documentation. The paper discusses methods of organizing work using both cloud and in-loop LLMs. A prototype of a tool for orchestrating heterogeneous models with access to the company's closed loop via VPN channels has been developed. A comparison of the solution of the problem by the intelligent assistant of the expert in the cloud and intra-contour modes is carried out. The developed prototype of a system operating without connection to external networks, providing a high level of confidentiality and import substitution, is described. The prototype provides for the separation of the roles of the subject user (expert) and the administrator, providing the possibility of secure interaction with the system. The developed intelligent assistant implements advanced contextual search and analytics/examination of project documentation based on information from specialized knowledge domains of limited access based on the use of RAG technology. The solution was implemented on the basis of a digital test site of the Higher School of Engineering MEPhI and is designed for use for educational tasks, educational and practical cases of industrial partners and for performing work in the interests of customers from the production sector, primarily Rosatom enterprises (including, but not limited to).