Design and Implementation of a High-Accuracy Text-to-SQL Pipeline with LLMs

Large Language Models (LLMs) are increasingly being explored for the Text-to-SQL task, with numerous techniques proposed by researchers to enhance their performance in SQL generation. In this thesis, we aim to design a high-accuracy Text-to-SQL pipeline suitable for real-world industrial applications, grounded in standard research practices. We conduct a comparative study of foundational approaches, evaluating alternative methods for Text-to-SQL such as LLM tool calling and prompting with linearized tables against conventional LLM prompting. Based on these findings, we systematically design an optimized pipeline that integrates state-of-the-art techniques, including M-Schema representation, advanced prompting strategies, schema linking, and validation. Furthermore, we investigate the use of reasoning-based models by comparing several frontier reasoning LLMs to identify the most effective one for the Text-to-SQL task. We then evaluate this model against non-reasoning models under identical conditions to determine the most effective overall approach for our use case. Following a comprehensive study, comparison, and evaluation of multiple existing approaches, we obtain an 'at least one column match' accuracy of 80\% on a 200-pair subset of the BIRD dataset. This work demonstrates the practical value of applying Text-to-SQL techniques in real industrial settings, contributing to a deeper understanding of how such solutions can be effectively engineered and deployed.