Využití velkých jazykových modelů jako chytré nápovědy

Abstract

This thesis focuses on the use of large language models as smart assistants in a domain-specific context. The goal was to evaluate the deployment and fine-tuning capabilities of open-source models Llama 3.2 and Gemma 3 using a dataset derived from official Debian manuals. The work includes an overview of current LLMs, their architectures, and tools for efficient customization. In the practical part, a web application was developed using a REST API to test and compare model responses. Emphasis was placed on real-world usability, computational performance, and output quality. The thesis concludes with an evaluation of the benefits and limitations of the tested models and a discussion of ethical and legislative aspects related to AI usage.