AgentNLQ: A General-Purpose Agent for Natural Language to SQL

2026-05-21 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

AgentNLQ, a new multi-agent method for Natural Language to SQL (NL2SQL) conversion developed by JPMorganChase, achieves 78.1% semantic accuracy on the BIRD benchmark. This solution addresses the challenge of generating accurate SQL queries from natural language questions, a task where LLMs often fall short compared to human experts. AgentNLQ integrates an optimized multi-agent orchestrator that plans, reflects, and self-corrects, an advanced schema enrichment method for context-aware metadata, and a multi-model agent configuration utilizing OpenAI GPT-4o and Anthropic Claude Opus 4.1. The system improved from a 60.2% baseline by incorporating orchestration and vector search (68-72%), then multi-model agents (76.4%), and finally its custom orchestrator with structured context compression (78.1%). It also reduced token usage by approximately 19.8% on the BIRD-bench financial dataset. Functional guardrails prevent DML operations, ensuring database integrity.

Key takeaway

For AI Engineers building enterprise NL2SQL solutions, you should prioritize multi-agent architectures that incorporate execution-grounded feedback and automated schema enrichment. Implementing a dual-ledger orchestrator, like AgentNLQ's, can significantly boost semantic accuracy to 78.1% while managing latency and token costs. Ensure your system includes functional guardrails to prevent unintended database modifications. Consider using multi-model agents for specialized tasks to maximize performance.

Key insights

Multi-agent orchestration with enriched schema and execution feedback significantly improves NL2SQL accuracy and efficiency.

Principles

• Combine fast (System 1) and slow (System 2) reasoning loops.
• Ground LLM feedback in direct SQL execution results.
• Automate schema enrichment for context-aware metadata.

Method

AgentNLQ uses an offline pipeline for metadata generation and an inference stage with a multi-agent orchestrator, SQL generator, and executor, employing a reason-generate-evaluate-replan loop with dual ledgers.

In practice

• Implement functional guardrails to prevent DML operations in NL2SQL systems.
• Use vector search for schema elements when context length is exceeded.
• Employ structured context compression to manage long conversation histories.

Topics

• NL2SQL
• Multi-Agent Systems
• LLM Orchestration
• Schema Enrichment
• BIRD Benchmark
• SQL Generation
• Context Management

Code references

• gkamradt/LLMTest_NeedleInAHaystack

Best for: Research Scientist, AI Architect, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Why I Stopped Using One Agent and Built a Multi-Agent Pipeline Instead — Single LLM agents struggle with complex, multi-faceted tasks by attempting too many competing mental models in one context.
• How We Built an AI Agent for Self-Service Analytics — Effective NL2SQL for business analytics requires structured data and a semantic layer, not just connecting an LLMs to a database.
• Building Cost-Efficient Agentic RAG on Long-Text Documents in SQL Tables — An Agentic RAG architecture can integrate LLMs with existing SQL databases for hybrid structured and semantic queries.
• Build AI-powered dashboard automation agents with NLP on Amazon Bedrock AgentCore — Multi-agent AI systems can automate complex, multi-step operational workflows by delegating tasks to specialized agents.
• AgentForge: Execution-Grounded Multi-Agent LLM Framework for Autonomous Software Engineering — Execution-grounded verification and specialized multi-agent decomposition significantly enhance LLM performance in autonomous software engineering.
• TAI #196: Quiet but Significant Agent Upgrades to Codex (Subagents) and Claude (Context) — Parallel agentic workflows and robust long-context processing are key to advancing AI development and deployment.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.