BudgetSchemaBench: A Budget-Swept Diagnostic for Schema Context in Text-to-SQL
BudgetSchemaBench tests how database schemas affect AI agents when constrained by token limits. It sweeps schema context budgets across a large catalog to see if accuracy improves with more or less schema information. The study shows execution accuracy is sensitive to this budget, especially under coverage-limited retrieval, and reveals a 'schema-memorization floor' when tables are removed.