04595ntm a22005777i 4500 000723868 CZ-PrVSE 20250629141137.0 m d cr n|||||||||| 250629s2025 xr fsbm 000 0 eng d NEZPRACOVANÝ IMPORT ABA006 cze ABA006 ABA006 rda Pivoda, Šimon ISIS:170048 dis Inovace velkých jazykových modelů (LLM) pomocí doladěné metody Retrieval-Augmented Generation (RAG) : Návrh a vyhodnocení eng Innovating Large Language Models (LLMs) with Fine-Tuned Retrieval-Augmented Generation (RAG) : Design and Evaluation Šimon Pivoda 2025 ?? stran : digital, PDF soubor Vedoucí práce: Richard Antonín Novák Bakalářská práce (Bc.)—Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky, 2025 Obsahuje bibliografii Textový (vysokoškolská kvalifikační práce) Rok obhajoby 2025 Artificial intelligence (AI) is reshaping the way organizations manage and interact with information. Its influence is visible not only in daily operations but also in critical strategic decision-making. In recent years, an innovative approach known as Retrieval-Augmented Generation (RAG) has emerged, extending the capabilities of large language models (LLMs) by providing access to relevant and up-to-date data sources. This enables companies to better leverage their internal knowledge in real time and apply AI across a wide range of use cases. However, despite the fast advancements in large language models, their direct deployment within corporate environments often faces major challenges. Main among these are concerns related to data security and the risk of data leakage. As a result, many organizations are hesitant to expose sensitive internal data to language models. Additionally, LLMs often suffer from response accuracy issues, including the generation of incorrect or hallucinated responses when lacking domain-specific knowledge. The RAG method capably furnishes resolutions for these challenges. Safe as well as governed admittance to internal information that is stored within specialized repositories known as vector databases is permitted instead of mainly depending on closed training datasets. This approach greatly curtails the risk regarding data breaches and inaccuracies. This paper digs into the basic tenets and salient constituents of the RAG architecture. Many efficacious document segmentation and retrieval strategies are implemented via vector databases and embeddings. In the practical part of this thesis, a complex RAG system is proposed and tested, targeting automotive data. A PostgreSQL database enhanced with vector search support (PGVector) is used to store and manage vector representations. The primary data source is the large dataset from the Kaggle platform. System evaluation is conducted using a combination of expert queries and ground truth. The evaluation aims to assess the system’s accuracy, output validation, and its ability to minimize hallucinations. The results indicate that the RAG-based architecture is effective in enhancing conversational systems with internal data sources, resulting in more accurate responses, increased reliability, and a notable reduction in hallucinated or fabricated content. Způsob přístupu: Internet aplikovaná informatika [obor bakal. práce] bakalářské práce fd132403 czenas bachelor's theses eczenas model evaluation prompt engineering embedding LLM RAG retriever AI context data Generative AI Novák, Richard, 1973- ntka173578 ths Korčák, Jiří, 1994- xx0336605 opn Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky kn20010709399 dgg https://insis.vse.cz/zp/91036/podrobnosti VŠKP v InSIS https://insis.vse.cz/zp/91036 Hlavní práce https://insis.vse.cz/zp/91036/posudek/vedouci Hodnocení vedoucího https://insis.vse.cz/zp/91036/posudek/oponent/85620 Oponentura https://insis.vse.cz/zp/91036/podrobnosti dc:identifier NEPOSILAT VSKP vse91036 250625 91036