A new paper says the classic Text-to-SQL trick of filtering database schemas before querying is now optional - and skipping it wins.
The study, posted to arXiv as "The Death of Schema Linking? Text-to-SQL in the Age of Well-Reasoned Language Models" (arXiv:2408.07702), revisits schema linking, the step where a Text-to-SQL pipeline narrows a database down to just the tables and columns relevant to a user's question before generating SQL. The authors found that newer large language models can pick out the right schema elements on their own, even when the full, uncurated schema is dumped into the context window alongside plenty of irrelevant tables. So their pipeline skips schema linking entirely whenever the schema fits in context, and instead leans on augmentation, selection, and correction steps to sharpen accuracy. That approach placed first on the BIRD benchmark, a standard test for Text-to-SQL systems, with 71.83% accuracy.
Schema linking was built to solve a real problem: LLMs used to get confused and drop required columns when a database had hundreds of tables. If that problem is fading as models get better at reasoning over noisy context, it simplifies the pipelines companies use to let non-technical staff query databases in plain English - fewer moving parts, fewer places for silent errors to creep in. It also points to a broader pattern: engineering workarounds built for weaker models can become dead weight once the underlying model improves.
Worth remembering, though - a 71.83% top score on BIRD also means Text-to-SQL still gets roughly three in ten queries wrong, schema linking or not.