FlexSQL: Flexible Exploration and Execution Make Better Text-to-SQL Agents

2026-05-04 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

FlexSQL is a novel text-to-SQL agent designed to improve performance over large analytical databases by enabling flexible interaction with the database throughout the reasoning process. Unlike traditional fixed-pipeline systems that retrieve schema elements once, FlexSQL can explore schema structures, inspect data values, and run verification queries at any stage. It generates diverse execution plans, implements them using either SQL or Python, and incorporates a two-tiered repair mechanism for error recovery. Using gpt-oss-120b on Spider2-Snow, FlexSQL achieved a 65.4% score, surpassing open-source baselines utilizing larger models like gpt-o3 and DeepSeek-R1. When integrated into Claude Code, it demonstrated over 10% relative improvement on Spider2-Snow, with flexible exploration and execution identified as key contributors to its effectiveness.

Key takeaway

For AI Engineers developing text-to-SQL agents for complex analytical databases, consider adopting a flexible interaction paradigm like FlexSQL. Your systems will benefit from dynamic schema exploration and multi-tiered error recovery, potentially outperforming models with more parameters by allowing agents to adapt and correct mistakes throughout the query generation process. Prioritize flexible execution and exploration in your next-generation agent designs.

Key insights

Flexible database interaction and execution planning significantly enhance text-to-SQL agent performance.

Principles

• Flexible database interaction improves error recovery.
• Diverse execution plans cover multiple interpretations.
• Two-tiered repair mechanisms enhance robustness.

Method

FlexSQL explores schema, inspects data, and runs verification queries dynamically. It generates diverse plans, implements them in SQL or Python, and uses a two-tiered repair for code and plan revisions.

In practice

• Integrate dynamic schema exploration in text-to-SQL.
• Employ multi-plan generation for query ambiguity.
• Utilize SQL/Python for task-specific execution.

Topics

• FlexSQL
• Text-to-SQL Agents
• Flexible Database Interaction
• Schema Exploration
• Execution Plan Generation

Code references

• StringNLPLAB/FlexSQL

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

Related on AIssential

• SQL Is Not Dead. Your AI Demo Is Just Hiding It. — AI natural language interfaces for databases primarily translate user input into SQL, not replace it.
• SAG: SQL-Retrieval Augmented Generation with Query-Time Dynamic Hyperedges — SAG dynamically links knowledge chunks using SQL joins at query time, enhancing RAG for structured data and multi-hop reasoning.
• One Flexible Tool Beats a Hundred Dedicated Ones — Flexible CLI tools outperform dedicated MCP tools for LLM agents due to improved compositionality and reduced context bloat.
• Oracle converges the AI data stack to give enterprise agents a single version of truth — Oracle's converged database architecture aims to solve data fragmentation and consistency issues for enterprise AI agents.
• SIRIUS-SQL: Anchoring Multi-Candidate Text-to-SQL in Execution Feedback — Multi-candidate Text-to-SQL reliability improves by diversifying generation, targeting repairs, and hybrid selection.
• Bridging the Gap Between Business Intelligence and Data Access — NL2SQL systems can democratize data access by translating natural language into executable SQL queries, enhancing decision-making speed.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.