LLM-POWERED DATABASE MIGRATION: A FRAMEWORK FOR KNOWLEDGE GRAPH SYSTEM EVOLUTION

Abstract

Database migration, particularly the translation of query languages, remains a significant barrier to mod ernizing data infrastructure. This challenge is especially acute as organizations adopt advanced knowledge graph (KG) technologies to support demanding applications in domains like smart cities and eHealth. This paper introduces a novel, LLM-powered framework for automated query translation, demonstrated through KG migration from RDF/SPARQL to LPG/Cypher. Our method leverages in-context learning with strategic exemplar selection and iterative refinement, achieving up to 89.6% translation accuracy and a 97.3% executable rate without requiring large parallel corpora or manual rule creation. Experiments on both the KQA Pro and enterprise-scale DBLP-QuAD datasets validate the approach’s effectiveness and scalability. With migration costs under $1.50 for thousands of queries, our framework offers an economically viable solution that reduces migration costs and accelerates the adoption of modern database technologies for next-generation applications.