Розробка семантичного метода поділу тексту для RAG систем, з керованим порогом та розміром ковзного вікна

The object of this study is Retrieval-Augmented Generation (RAG) systems used to improve the quality of responses by large language models (LLMs). The task addressed is to improve the efficiency of the semantic text segmentation stage in such systems, which directly affects the accuracy of extracting relevant fragments. The work reports a method of semantic text segmentation for RAG systems, based on the sliding window technique with a dynamically changing size. The method devised uses embedding models and makes it possible to take into account the semantic context of the text. The adjustable value of the cosine similarity threshold used in semantic splitting makes it possible to additionally increase the relevance of query formation to LLM. The developed algorithm for setting this threshold value makes it possible to more fully take into account the specificity of the query subject. Compared to advanced methods of semantic text segmentation, the method devised provides the following gains depending on the maximum document size parameter: IoU from 0.2 % to 2.8 %, precision from 0.4 % to 3.1 %, omega precision from 1.4 % to 14.8 %. The gains are primarily associated with text processing at the level of semantically complete units in the form of sentences, rather than tokens. In addition, the dynamic sliding window technique allowed for better adaptation to the text structure. The results are valid within the framework of the used evaluation, which covers heterogeneous text datasets, and could be applied in practice when building RAG systems in industries with high requirements for preserving the semantic integrity of the text, for example, in law, science, or technology. The algorithms that implement the proposed method are posted on GitHub as Python libraries