Development of an intelligent conversational assistant with dynamic selection of neural language models and a knowledge retrieval mechanism

Abstract

This paper presents the development of an intelligent conversational assistant designed to support users in obtaining and clarifying information about university class schedules. The evolution of dialogue systems is considered, from early rule-based text dialogue programs based on pattern matching of input utterances to modern virtual assistants implemented using neural transformer architectures and large language models. The relevance of the study is determined by the use of large language models in combination with Retrieval-Augmented Generation (RAG), which makes it possible to generate responses based on information retrieved from external sources. An architecture of an intelligent conversational assistant has been developed that integrates a RAG mechanism, dynamic routing between several neural language models, and automatic updating of the knowledge base through parsing Admiral F. F. Ushakov State Maritime University web resources. The proposed approach ensures robust processing of variable user queries and response generation based on the retrieved context. Unlike existing solutions that rely on a fixed language model, the proposed architecture implements adaptive selection of a generative model depending on the type of user query. A method for domain-oriented context construction for processing dynamically updated tabular schedule data has been developed, including specialized web page parsing, semantic structuring of the extracted information, and prompt engineering rules. To improve the factual reliability of responses, a mechanism is proposed that constrains generation to the retrieved data and applies terminological filtering. An ETL pipeline for updating the knowledge base has been implemented, enabling automatic updating of information about university class schedules. A comparative study of the effectiveness of the Qwen, GigaChat, and DeepSeek language models in processing user queries within the considered subject domain has been conducted. The obtained results demonstrate improved response accuracy and confirm the effectiveness of the proposed conversational assistant architecture.