Institutional Repository of Vilnius University
Pažangių paieškos metodų taikymas lietuviškų dokumentų RAG sistemose /
Abstract
dc:descriptionThis thesis presents a study on the effectiveness of various retrieval methods for Lithuanian documents in RAG (Retrieval Augmented Generation) systems. With the increasing popularity of large language models (LLM) in various information systems, there is a challenge to effectively integrate these models with documents in specific languages such as Lithuanian. One of the main challenges is the morphological complexity of the Lithuanian language, which complicates effective semantic document retrieval. Therefore, the aim of this work is to investigate and compare various retrieval methods for Lithuanian document RAG systems to determine their effectiveness and applicability. The thesis provides an analysis of the theoretical foundations of RAG systems, reviews the specific challenges of the Lithuanian language, and examines the principles of various retrieval methods. The research methodology section describes data sets, evaluation metrics, and experimental methodology. The retrieval methods chapter thoroughly analyzes seven different methods: Top-K, sentence window, auto-merging, hybrid with reranking, maximum marginal relevance, hypothetical document embeddings, and hypothetical query embeddings. The results analysis section provides a comparison of the methods based on semantic answer similarity, recall score, and the number of LLM calls. The thesis concludes with findings and recommendations regarding the optimal retrieval methods for Lithuanian document RAG systems.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mackonis, Dovydas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD239024852&prefLang=en_US