Building an Agentic System for Question-Answering of Public Administrative Documents and Procedures in Vietnamese

Abstract

Public administrative procedures in Vietnam involve complex documentation requirements that citizens must navigate to access government services. The complexity and language-specific nature of Vietnamese administrative documents pose significant challenges for automated question-answering systems. This paper presents an agentic system architecture for intelligent question-answering over Vietnamese administrative documents and procedures. Our system leverages large language models (LLMs) with retrieval-augmented generation (RAG), multi-agent orchestration, and specialized Vietnamese natural language processing techniques. The system employs a hierarchical agent architecture in which specialized agents handle document retrieval, semantic understanding, procedure extraction, and response generation. We evaluate our approach on a curated dataset of Vietnamese administrative documents covering various domains, including citizenship, residency, business registration, and social services. The experimental results demonstrate significant improvements in accuracy and relevance compared to the baseline approaches, achieving 87.3% accuracy on factual questions and 82.1% on procedural queries. The system's agentic architecture enables dynamic reasoning, multi-step procedure guidance, and context-aware responses that adapt to user queries.