Boost Text-to-SQL Accuracy with Dynamic Few-Shot Learning and Alignment

From June 22‑27, 2025, SIGMOD 2025 was held in Berlin, gathering top scholars to discuss database systems, big data, and data analytics.

Alibaba Cloud AI Search team’s paper “OpenSearch‑SQL: Enhancing Text‑to‑SQL with Dynamic Few‑shot and Consistency Alignment” was accepted. The work introduces a self‑taught dynamic few‑shot mechanism and a consistency‑alignment strategy to improve NL2SQL performance.

The paper highlights the challenges of Text‑to‑SQL in the LLM era, such as unclear architecture, under‑utilization of large‑model capabilities, and over‑fitting in fine‑tuning. Its core goal is to boost query‑level accuracy of Text‑to‑SQL by building a complete pipeline that addresses missing database knowledge, model hallucinations, insufficient few‑shot information, and error accumulation in multi‑agent collaboration.

Key Components of OpenSearch‑SQL

A robust pipeline covering preprocessing, information extraction, generation, and error correction.

Self‑taught Few‑shot: transforms simple query‑SQL pairs into Query‑CoT‑SQL via a self‑learning mechanism.

Alignment mechanism for heterogeneous agents to reduce information loss and hallucination.

Paper Details

Title: OpenSearch‑SQL: Enhancing Text‑to‑SQL with Dynamic Few‑shot and Consistency Alignment

Authors: Xiangjin Xie, Guangwei Xu, Lingyan Zhao, Ruijie Guo

Link: https://arxiv.org/abs/2502.14913

Presentation at SIGMOD 2025

Alibaba Cloud delivered a keynote titled “Free‑Flow Search: Trends in Alibaba Cloud AI Search Algorithms”, discussing AI‑driven search advancements, industry‑specific models, multimodal search, NL2Search, and WebSearch, aiming to improve search efficiency and accuracy across applications.

For implementation details, refer to the Alibaba Cloud AI Search Open Platform documentation: NL2SQL service configuration (https://x.sm.cn/3BXRXQs).