SQL-ACT : content-based and history-aware input prediction for non-trivial SQL queries

Abstract

This thesis presents SqlAct, a SQL auto-completion system that uses content-based and history-aware input prediction to assist in the process of composing non-trivial queries. By offering the most relevant suggestions to complete the partially typed query first at the word-level and then at the statement-level, SqlAct hopes to help both novice and expert SQL developers to increase their productivity. Two approaches are explored: word-level suggestions are optimized based on the database's schema and content statistics, and statement-level suggestions that rely on Long Short-term Memory (LSTM) Recurrent Neural Networks language models trained on historical queries. The word-level model is integrated in a responsive command-line interface database client which is evaluated quantitatively and qualitatively. Results shows SqlAct provides a highly-responsive interface that makes high quality suggestions to complete the currently typed query. Possible directions for integration with the word-level model in the command-line tool are explored as well as the planned evaluation techniques.