Contrastive Self-Refinement for Low-Cost Adaptation in Real-World Text-to-SQL

Deploying Text-to-SQL systems in real-world industries faces a "last-mile" bottleneck: the high maintenance cost of adapting to frequent schema changes and the lack of domain-specific knowledge in raw databases. Traditional fine-tuning methods struggle to keep pace with evolving business logic, while standard in-context learning lacks the robustness to bridge the semantic knowledge gap. To address these challenges, we propose SEEK-SQL, a training-free, self-evolutionary multi-agent framework designed for low-cost adaptation. Our framework introduces a novel Contrastive Self-Refinement (CRF) strategy that mimics human learning: by contrasting positive and negative execution trajectories, the system generates reusable guidelines to "learn" new schema rules on the fly without parameter updates. Additionally, we integrate a Sniffer agent to dynamically retrieve external knowledge, ensuring robustness against ambiguous queries. Experiments on Spider and BIRD show that SEEK-SQL achieves state-of-the-art performance among in-context learning methods while reducing token consumption by significantly compared to other multi-agent frameworks. These results demonstrate SEEK-SQL's potential as a cost-efficient, agile solution for industrial Text-to-SQL deployment.