AgentNLQ: A Multi-Agent Framework for Natural Language to SQL Conversion

Forum topic · 小凯 · 2026-05-21

Summary

This paper presents AgentNLQ, a multi-agent method for Natural Language to SQL (NL2SQL) conversion that addresses the gap between LLM-generated queries and expert-written SQL. The system achieves 78.1% semantic accuracy on the BIg Bench for LaRge-scale Database (BIRD) benchmark. The approach leverages a semantically enriched representation of the user-provided schema, incorporates user-provided business rules, and produces accurate SQL queries through an optimized orchestrator that plans, orchestrates, reflects, and self-corrects using LLMs. The authors also introduce an advanced schema enrichment method that creates context-aware metadata to improve accuracy, and demonstrate generalizability across multiple domains and datasets. The work highlights how multi-agent orchestration combined with schema enrichment can advance NL2SQL reliability for enterprise and research applications involving relational databases.

Paper Overview

Research field: cs.AI
Authors: Olena Bogdanov, Yeunji Jung, Chandra Dhir
Release date: 2026-05-21
arXiv: 2505.01254

Abstract

Natural language to SQL (NL2SQL) conversion is an important problem for researchers and enterprises due to the ubiquitous importance of relational databases in broad-ranging practical problems. Despite the rapid advancements in the capabilities of LLMs, NL2SQL has not reached parity in accuracy with human expert SQL writers, hence needing additional improvements in NL2SQL algorithms.

This study presents a new multi-agent method for NL2SQL that achieves 78.1% semantic accuracy on the BIg Bench for LaRge-scale Database (BIRD) benchmark. The method leverages a semantically enriched representation of the user-provided schema, adds user-provided business rules, and produces accurate SQL queries.

Key Contributions

(a) Optimized orchestrator: Designed a new orchestrator in a multi-agent solution that uses LLMs to plan, orchestrate, reflect, and self-correct in order to generate accurate SQL queries.
(b) Advanced schema enrichment: Developed a method that creates context-aware metadata to improve query-generation accuracy.
(c) Empirical validation: Demonstrated accuracy and generalizability across different domains and datasets by evaluating the method on the BIRD-SQL benchmark, reaching 78.1% semantic accuracy.

Significance

The work shows that combining multi-agent orchestration, self-correction, and rich schema metadata can meaningfully close the gap between LLM-generated NL2SQL output and expert human SQL, with practical implications for enterprise analytics and research workflows involving relational data.