Scholarlink

This paper investigates how the model size affects the ability of a Generative AI Language Model, or briefly a GLM, to support the text-to-SQL task for databases with large schemas typical of real-world applications. The paper first introduces a text-to-SQL framework that combines a prompt strategy and a Retrieval-Augmented Generation (RAG) technique, leaving as flexibilization points the GLM and the database. Then, it describes a benchmark based on an open-source database featuring a schema much larger than the schemas of most of the databases in familiar text-to-SQL benchmarks. The paper proceeds with experiments to assess the performance of the text-to-SQL framework instantiated with the benchmark database and GLMs of different sizes. The paper concludes with recommendations to help select which GLM size is appropriate for a text-to-SQL scenario, characterized by the difficulty of the expected NL questions and the data privacy requirements, among other characteristics.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Small, Medium, and Large Language Models for Text-to-SQL

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Aiko Oliveira,
  • Eduardo Nascimento,
  • João Pinheiro,
  • Caio Viktor S. Avila,
  • Gustavo Coelho,
  • Lucas Feijó,
  • Yenier Izquierdo,
  • Grettel García,
  • Luiz André P. Paes Leme,
  • Melissa Lemos,
  • Marco A. Casanova

摘要

This paper investigates how the model size affects the ability of a Generative AI Language Model, or briefly a GLM, to support the text-to-SQL task for databases with large schemas typical of real-world applications. The paper first introduces a text-to-SQL framework that combines a prompt strategy and a Retrieval-Augmented Generation (RAG) technique, leaving as flexibilization points the GLM and the database. Then, it describes a benchmark based on an open-source database featuring a schema much larger than the schemas of most of the databases in familiar text-to-SQL benchmarks. The paper proceeds with experiments to assess the performance of the text-to-SQL framework instantiated with the benchmark database and GLMs of different sizes. The paper concludes with recommendations to help select which GLM size is appropriate for a text-to-SQL scenario, characterized by the difficulty of the expected NL questions and the data privacy requirements, among other characteristics.