Functional Architecture of an Artificial Intelligence-Based Virtual Tax Assistant System

Abstract

This paper presents the development of the functional architecture
of an artificial intelligence-based virtual tax assistant system designed to operate on
the basis of the national tax legislation. The proposed system is built upon a Large
Language Model (LLM), the Retrieval-Augmented Generation (RAG) framework,
and a vector database. To capture the semantic relationships among tax legislation
documents, a knowledge graph was constructed, and a graph-based semantic
chunking approach was employed. Furthermore, the document retrieval and context
generation mechanisms were enhanced using the Graph-RAG approach.
Experimental results demonstrate that the proposed method improves the accuracy
of generating contextually relevant responses compared with the conventional RAG
algorithm. The proposed architecture enhances the effectiveness of the virtual tax
assistant and enables the provision of reliable advisory services based on the
national tax legislation