М. В. Гончаров The paper comprises the findings of the integrated study of using large language models (LLMs) artificial intelligence (AI) technologies in the national library information practice. The comparative analysis is accomplished for Russian (GigaChat, YaGPT) and foreign (ChatGPT, Claude, LLaMA, Mistral, DeepSeek) LLM applications for three key library tasks, i. e. semantic deconstruction of user queries, recognition of handwritten catalog cards, and automated correction of text errors. The system testing on representative query selection, handwritten cards and fulltext document images via APIs was accomplished. The critical limitations of existing solutions are revealed: instability and degradation of models, excessive censorship with high percentage of false triggering, inexistent data generation (hallucinations), and unpredictability of structured inference, or cultural and linguistic barriers. Based on the study findings, the specialized IRBIS AI system based on the Mixture of Experts architecture was designed to ensure prompt and stable processing of bibliographic data. The authors discuss the practical implementation in J-IRBIS 2.0 module with three modes of AI-support: reference services, query semantic processing, and intellectual literature selection. The prospects for building the library portal with AI-controlled interface, cataloguing automation through multimodal models, integrated processing of unstructured data, and image catalogs transformation, are outlined.
Научная статья 
справочное обслуживание