DecoSearch: Complexity-Aware Routing and Plan-Level Repair for Text-to-SQL

2026-06-16 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

DecoSearch is a novel training-free framework designed to enhance Large Language Models' (LLMs) performance in translating natural language to SQL, particularly for complex queries requiring multi-step, data-aware reasoning. It employs a lightweight Schema Selector to prune database schemas and an LLM Judger to route questions, either directly generating SQL for simple queries or escalating complex ones to a Directed Acyclic Graph (DAG) of sub-questions. A RAG component grounds the decomposer, and a Topology Refiner repairs flawed reasoning plans. DecoSearch achieves 70.53% execution accuracy on BIRD and 88.31% on Spider with a DeepSeek backbone, consuming significantly fewer tokens than competing methods, and functions as a model-agnostic wrapper to improve existing SQL generation backbones.

Key takeaway

For Machine Learning Engineers or NLP Engineers building Text-to-SQL solutions, DecoSearch offers a compelling training-free approach to significantly improve accuracy on complex queries. You should consider integrating this framework as a model-agnostic wrapper to enhance your existing fine-tuned SQL generation backbones, potentially reducing token consumption by an order of magnitude while achieving higher execution accuracy on benchmarks like BIRD and Spider.

Key insights

DecoSearch improves Text-to-SQL accuracy for complex queries via complexity-aware routing and plan-level repair.

Principles

• Route queries by complexity for optimal reasoning effort.
• Decompose complex Text-to-SQL questions into sub-questions.
• Repair flawed reasoning plans at the topology level.

Method

DecoSearch prunes schema, judges query complexity, decomposes complex queries into DAGs of sub-questions, grounds with RAG, and refines the plan if execution fails.

In practice

• Implement schema pruning to reduce LLM context.
• Break down complex natural language queries into atomic sub-questions.
• Integrate RAG to provide relevant training examples for decomposition.

Topics

• Text-to-SQL
• Large Language Models
• Query Decomposition
• Schema Pruning
• Retrieval-Augmented Generation
• Execution Accuracy

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

Related on AIssential

• QueryWeaver: Reliable Multi-Tool Query Execution Planning via LLM-Based Graph Generation — QueryWeaver transforms natural language into structured graphs for reliable multi-tool query execution using a deterministic, dependency-resolving planner.
• TAHOE: Text-to-SQL with Automated Hint Optimization from Experience — TAHOE optimizes Text-to-SQL LLM prompts by dynamically learning and applying error-driven hints from development and deployment experiences.
• Aligning Tree-Search Policies with Fixed Token Budgets in Test-Time Scaling of LLMs — Budget-Guided MCTS dynamically adjusts LLM tree-search from broad exploration to focused refinement based on remaining token budget.
• Executable Schema Contracts: From Automatic Ingestion to Multi-Source Retrieval — The system automatically derives an executable schema from multi-source data to unify knowledge graph construction and multi-tool query retrieval.
• Retrieval is Cheap, Show Me the Code: Executable Multi-Hop Reasoning for Retrieval-Augmented Generation — Reformulating multi-hop RAG as program synthesis and execution enhances verifiability and performance.
• We Need SQL for AI Programs (And IBM Might Have Built It) — Span Queries propose a declarative optimization layer for GenAI programs, akin to SQL for databases.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.