Návrh metodiky pro automatickou klasifikaci dokumentů ve veřejné správě s využitím velkých jazykových modelů

This thesis deals with the design of a system for automatic document classification in public administration and its verification according to the requirements of the European regulatory framework. The design is based on an analysis of the document repository of the Ministry of Industry and Trade and an evaluation of state-of-the-art classification methods. The motivation is the need for efficient sorting of millions of documents, which is key to preventing data leaks and the future development of intelligent services. Analysis of current approaches and market consultations have shown that modern language models provide the most suitable solution. Based on this, the Retrieval Augmented Classification method was chosen, supplemented by a zero-shot procedure for general document types. The thesis also includes validation of the design according to the areas defined in the EU AI Act, which shows that the system is compatible with the requirements for data governance, documentation, oversight, accuracy, and cybersecurity. The limitations of the work lie mainly in the absence of representative data for practical verification of the design. However, the resulting concept represents a feasible and transferable basis for pilot testing and further contributes to the discussion on the safe use of artificial intelligence in public administration.