DEVELOPMENT OF AN AI-BASED LEGAL CONSULTATION SYSTEM USING HIERARCHICAL RAG, HYBRID RETRIEVAL, AND LARGE LANGUAGE MODELS FOR VIETNAMESE LEGAL DOCUMENT SEARCH

Abstract

Accessing accurate Vietnamese legal information remains challenging due to the vast number of regulations, complex hierarchical structures, and the limitations of traditional keyword-based search tools. This thesis develops LegalMate, an AIpowered legal assistant that addresses these issues through a novel 3-tier RetrievalAugmented Generation (RAG) architecture combined with hybrid retrieval (BM25 + dense vector search), query enhancement, and cross-encoder reranking. The system was implemented using a Python/FastAPI backend with PostgreSQL + pgvector for efficient hierarchical vector search, Google Gemini-1.5- flash for response generation and query enhancement, and a cross-platform Flutter mobile application (Android/iOS). Evaluation on a diverse legal question dataset shows that the proposed 3-tier RAG approach achieves recall@5 ≥ 80%, precision@5 ≥ 60%, groundedness ≥ 4.0/5.0, and an average response time of 8–20 seconds— significantly outperforming baseline keyword and single-tier vector search methods. The results demonstrate the effectiveness and practical feasibility of structured multi-tier RAG for the Vietnamese legal domain, providing a reliable, user-friendly, and source-cited legal assistant suitable for both citizens and professionals.