Multi-Agent Orchestration for Enhanced Text-to-SQL Generation: A Schema-Aware Approach with Self-Correction

Abstract: The unification of the functionality of databases and natural language processing Databases has traditionally only provided a means to be able to write SQL queries, however there is an ongoing challenge of converting NL (Natural Language) queries into Structured Query Language(SQL). However, LLMs have significant challenges in handling complex schema structures and long-range dependencies, as well as a tendency toward structural hallucination. This paper presents a novel multi-agent architecture based on LangGraph with the aim to close the gap between NL and SQL through specialized decomposed reasoning. Our implemented system incorporates a robust AST-based Self-Correction loop, a query planning agent for chain-of-thought reasoning, and a semantic-aware Schema Linking Agent using Sentence Transformers. Experimental results on 100 examples from the Spider benchmark show that the complete four-phase system (Model C) achieves the highest Execution Accuracy at 76.0%, outperforming the monolithic zero-shot baseline (63.0%) by +13.0 percentage points. The Schema Linking Agent alone (Model A) achieves 73.0% EX, while the addition of the AST-based Self-Correction loop (Model B) reaches 75.0% EX with an average correction rate of 0.16. Semantic value mapping, dynamic schema discovery, and query complexity routing are identified as crucial paths for future enterprise deployment by a thorough gap analysis.

Keywords: Text-to-SQL, Multi-Agent Systems, Large Language Models, Schema Evolution, Self-Correction, Vector Search, Cross-Database Reasoning.