Development of an intellectualized program system based on rag-architecture

Abstract

The article presents an approach in the creation of an intellectualized dialogue system for automation of communication with university applicants during admission campaigns. Retrieval-Augmented Generation (RAG) architecture, which combines search of relevant information with generation of a response using a Large Language Model, is the basis of the development. Fine-tuning of the GPT-2 model with regulatory documentation of a higher educational facility was used to provide support for Ukrainian language. A hybrid mechanism of response generation, which combines fragments extracted from a knowledge base with segments generated by the language model, was proposed. An architecture, which ensures the relevance, accuracy and completeness of data, as well as lowering the workload on university employees, was realized. The quality of the system was rated using both objective (response speed, BERTScore, relevance) and subjective metrics (completeness, convenience, search flexibility), which allowed to record the efficiency of the approach. Presented instrument demonstrates the perspective of the usage of RAG approaches in applied tasks of the educational sphere, especially in support systems for admission committees.