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

DecoSearch is a new training-free framework for improving text-to-SQL translation, specifically targeting complex queries that require multi-step reasoning. It utilizes a lightweight Schema Selector to identify relevant database components, an LLM Judger to determine the need for query decomposition, and a Directed Acyclic Graph (DAG) to manage atomic sub-questions, achieving 70.53% accuracy on the BIRD dataset and 88.31% on Spider. This model-agnostic approach enhances existing SQL generation models without altering their architecture, significantly reducing token consumption compared to other methods, which is crucial for practitioners aiming for efficient and effective query processing.