DecoSearch Improves Text-to-SQL with Complexity-Aware Routing and Repair.

DecoSearch introduces a novel, training-free framework designed to significantly improve the accuracy and efficiency of text-to-SQL translation using Large Language Models (LLMs). This system intelligently routes natural language queries based on their complexity, ensuring that straightforward questions are processed directly while more intricate ones are decomposed into a Directed Acyclic Graph (DAG) of atomic sub-questions. Each sub-question is then solved through a targeted SQL generation step. A key innovation of DecoSearch is its multi-component architecture. It includes a lightweight Schema Selector to prune irrelevant database schema, an LLM Judger to determine query decomposition necessity, and a Retrieval-Augmented Generation (RAG) component to ground the decomposer with relevant examples. Furthermore, a Topology Refiner is integrated to restructure the reasoning plan when execution failures indicate a flawed decomposition rather than a simple SQL error. The framework has demonstrated state-of-the-art performance on challenging benchmarks such as BIRD and Spider, achieving high execution accuracy with significantly reduced token consumption compared to other methods. Its model-agnostic nature allows it to function as a wrapper, consistently enhancing existing fine-tuned SQL generation backbones without requiring any modifications to their internal pipelines.