Project Profile

AAP-SQL

在控制推論成本的前提下提升 Text-to-SQL 準確率的碩士論文研究專案。

Completed
Status
Completed
Portfolio
Featured
Role
Master's thesis research project
Read overview
Text-to-SQLLLMDatabasesEvaluation

AAP-SQL 是我的碩士論文研究專案，關注的是 Text-to-SQL 系統如何在複雜資料庫中補充真正有用的欄位、資料值與關係線索，同時控制大型語言模型的推論成本。

專案範圍

  • 以 BIRD 開發集研究跨資料庫的 Text-to-SQL 問題
  • 以 selective retrieval 補充核心欄位的真實資料值與關係
  • 以 execution accuracy、Token 用量與消融實驗檢查方法取捨

這個 Profile 保留研究問題、結果與限制的入口；完整方法說明放在相關 Overview Article。

What changed or was learned

可驗證成果

62.58% execution accuracy on BIRD

The reported local reproduction answered 960 of 1,534 development questions under a fixed Gemini 2.5 Flash-Lite comparison.

Accuracy with a bounded token budget

AAP-SQL used 3.32M total tokens, substantially below the two multi-stage comparison methods reported in the same experiment.

Traceable sources

Evidence Trail

  1. Overview Article

    Method, evaluation table, ablations, limitations, and the thesis context.

    Inspect source

Writing