Підвищення якості семантичного пошуку в базах знань із використанням векторних подань

Abstract

This paper presents an applied evaluation of semantic search quality in a knowledge-base retrieval system. The study uses a reproducible pipeline that combines preprocessing, chunking, vector indexing, querying, and benchmark based assessment across two domains. The system's performance was evaluated using a specially prepared set of queries and widely used metrics for assessing search effectiveness. Results demonstrate that contextual embeddings achieve superior ranking quality, while lexical methods ensure comprehensive coverage. The proposed approach supports iterative quality control and further development of semantic and hybrid retrieval systems for enterprise knowledge management.