A legal support system for entrepreneurs based on a Retrieval-Augmented Generation architecture
Egor S. Boykov, Lead Specialist of Infrastructure and System Administration Dept. The State Budgetary Institution of the Sverdlovsk Region “Electronic Government Operator”
Natalya S. Kolyeva, Cand. Sc. (Ped.), Acting Head, Associate Prof. of Information Technology and Statistics Dept. Ural State University of Economics
Liudmila V. Kortenko, Cand. Sc. (Econ.), Associate Prof. of Information Technology and Statistics Dept. Ural State University of Economics
Abstract
The aim of the study is to develop and experimentally validate a RAG system architecture for information and legal support of entrepreneurs when performing control and supervisory activities. The architecture is built upon hybrid search, which combines vector representation and the BM25 lexical algorithm via Reciprocal Rank Fusion, result reranking using a cross-encoder, metadata filtering across the regulatory act lifecycle, and a dual-loop structure incorporating expert verification. The experiment was conducted on a validated test set of 600 queries with reference answers. Compared to the baseline configuration, search recall (Recall@20) increased from 0.64 to 0.92, ranking quality (NDCG@10) from 0.48 to 0.87, faithfulness from 0.72 to 0.94, and citation accuracy from 0.55 to 0.91. The contribution of individual components was established through ablation analysis, and the robustness of the obtained results was further verified across five non-overlapping stratified subsamples. The methodologies used for constructing the test set and evaluating configurations ensure the reproducibility of the experiment within the Russian legal domain. The practical value of this work consists in reducing business labor costs in preparing for control and supervisory activities, as well as mitigating the risks of misinterpreting regulatory acts.
Keywords: control and supervision activities; regulatory compliance; artificial intelligence; RAG architecture; hybrid search; reranking
For citation: Boykov E.S., Kolyeva N.S., Kortenko L.V. A legal ssupport system for entrepreneurs based on a Retrieval-Augmented Generation architecture. Digital Models and Solutions. 2026. Vol. 5, no. 3, pp. 28–49. DOI: 10.29141/2949-477X-2026-5-3-3. EDN: HXTTUO.
