One Approach to Relational Database Search Using a Graph RAG System

Abstract

Relational databases remain the backbone of business information systems, but working with them requires knowledge of SQL. While Large Language Models have significantly brought closer the possibility of querying using natural language, turning those questions into reliable and precise answers from relational data remains a challenge. Pure textual RAG (Retrieval Augmented Generation) often ignores schemas, keys, and constraints, while the direct translation of a natural language query into SQL using LLMs is fragile and difficult to verify. This paper presents the GraphRAG (Graph Retrieval Augmented Generation) solution to the problem, which enables natural language access to relational data.