Multi-Agent NL2SQL Method Achieves 78.1% Accuracy on BIRD Benchmark

A new multi-agent method for natural language to SQL (NL2SQL) conversion has achieved 78.1% semantic accuracy on the BIRD benchmark, as detailed in a paper on arXiv. The approach uses a semantically enriched representation of user-provided schema and incorporates user-provided business rules to generate accurate SQL queries. Key contributions include an optimized orchestrator in a multi-agent solution that leverages LLMs for planning, orchestration, reflection, and self-correction, as well as advanced schema enrichment. The study addresses the persistent gap between LLM-based NL2SQL and human expert SQL writers, aiming to improve accuracy for practical applications relying on relational databases.