Text-to-SQL mapping process plays a crucial role in enabling non-technical users to interact with relational databases using natural language. While Large Language Models (LLMs) have shown promising results in benchmark datasets, their performance in real-world settings often deteriorates. In this work, we propose a modular and adaptive agent that leverages the opensource LLM DeepSeek to perform translations with integrated feedback and query optimization. Our agent is prompt-engineered to refine query generation based on user or system-provided corrections. Using the TPC-H and Mondial benchmarks, as well as a real-world database, we demonstrate improved accuracy and execution efficiency, highlighting the impact of feedback loops and heuristic-based query rewriting.
Petrola et al. (Mon,) studied this question.