Retrieval Augmented Generation for Ukrainian Government Services: A Comparative Evaluation of the Approaches

dc.contributor.advisorКурочкін, Андрійuk_UA
dc.contributor.authorМаринич, Антонuk_UA
dc.date.accessioned2025-09-05T06:53:49Z
dc.date.available2025-09-05T06:53:49Z
dc.date.issued2025
dc.description.abstractRetrieval Augmented Generation or RAG is a method that is used to improve the quality of retrieval for LLMs, to avoid hallucinations and be aware of all the changes in the data. This approach integrates LLMs with external data source by building a vector index. This thesis presents a comprehensive study on how different RAG approaches perform in Ukrainian Governmental Services domain. I establish a non-RAG baseline using GPT-4.1-mini model and iteratively perform tests on different configurations of RAG approaches. I have also created a dataset with 500 open questions about Ukrainian Governmental Services using GPT-o4-minihigh model. My best results comparing to the baseline are 13.25% improvement in LLM Judge Score using CRAG with Hypothetical Document Embedding and Reranking and 10% improvement on Factual Correctness using CRAG with Reranking. en_US
dc.identifier.urihttps://ekmair.ukma.edu.ua/handle/123456789/36460
dc.language.isoen_US en_US
dc.statusfirst publisheden_US
dc.subjectRetrieval Augmented Generation (RAG)en_US
dc.subjectLarge language models (LLMs)en_US
dc.subjecthallucinationsen_US
dc.subjectUkrainian Governmental Services domainen_US
dc.subjectbachelor`s thesisen_US
dc.titleRetrieval Augmented Generation for Ukrainian Government Services: A Comparative Evaluation of the Approaches en_US
dc.typeOther en_US
Files
Original bundle
Now showing 1 - 2 of 2
Loading...
Thumbnail Image
Name:
Marynych_Bakalavrska_robota.pdf
Size:
1.07 MB
Format:
Adobe Portable Document Format
Loading...
Thumbnail Image
Name:
Marynych_Bakalavrska_robota_1.pdf
Size:
672.59 KB
Format:
Adobe Portable Document Format
License bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
license.txt
Size:
1.71 KB
Format:
Item-specific license agreed upon to submission
Description: