New benchmark and adaptive embeddings boost SQL schema retrieval performance

Researchers have introduced a new benchmark and corpus-adaptive embeddings for SQL schema retrieval, a crucial step in text-to-SQL tasks that involves identifying relevant tables and columns within large databases. They adapted five existing text-to-SQL datasets to function as retrieval tasks and found that standard text and code embedders performed poorly. By fine-tuning a 305M-parameter embedder using synthesized queries and hard negatives, they significantly improved recall@10 from 60.4% to 75.6%, establishing schema linking as a distinct retrieval problem and demonstrating a practical method for its deployment at enterprise scale. AI

IMPACT Improves the efficiency and accuracy of text-to-SQL systems, potentially accelerating enterprise adoption of AI for data analysis.

RANK_REASON The cluster contains an academic paper detailing a new benchmark and methodology for SQL schema retrieval. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• BIRD
• LiveSQLBench
• SQL

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →