ExploreSQL: Grounding Arabic Text-to-SQL Generation Through Iterative Database Exploration

Abstract

ExploreSQL presents a new inference-time framework for Arabic Text-to-SQL generation that lets the model directly query the target SQLite database — issuing exploratory SELECT statements and observing real results — before committing to a final answer, rather than relying solely on a static, pre-retrieved schema context. Built on a newly cleaned 3,830-example Arabic evaluation corpus (deduplicated and execution-validated from Spider/AraSpider), the method is tested across three exploration depths (1–3 turns) using Devstral Medium. Results show that allowing even one extra exploration turn boosts execution accuracy from 64.56% to 69.10%, with diminishing returns beyond that, while surface-level metrics (Exact Match, BLEU, Token F1) stay stable — indicating exploration mainly helps the model verify data-specific details (like exact value spelling) rather than change query structure. The work extends the authors' earlier RAG-based Arabic Text-to-SQL pipeline by replacing static retrieval with real, executed database evidence.