AI Can Now Query Databases Without Being Told Where to Look

For years, when you wanted to ask a database a question in plain English, engineers had to build a middle step. Before the AI could write a database query, a separate program would guess which tables and columns were relevant and throw the rest away. This filtering step, called 'schema linking,' was supposed to help. But it often backfired: if the filter guessed wrong, the AI never saw the data it needed, and the answer came out wrong.

In this new research, a team tested the newest generation of large language models (AI trained on huge amounts of text) and found something surprising. These models are now good enough to sift through lots of irrelevant information on their own. So the team simply stopped filtering. Instead, they hand the AI the entire database structure and let it figure out what matters. They also added techniques like giving the AI examples, having it double-check its work, and correcting its own mistakes.

The result: their system now ranks first on BIRD, a widely used public test where computers must translate plain-English questions into the correct database queries. It scored 71.83% accuracy, beating previous approaches. That may sound modest, but this test is notoriously hard — it uses messy, real-world databases, not tidy textbook examples. Every percentage point represents many more questions answered correctly.

Why does this matter beyond research papers? Because 'text-to-SQL' — turning a plain question into a database search — is what powers many business tools. It is how a manager might one day ask, 'How many customers in Texas churned last quarter?' and get an instant chart. When the filtering step disappears, these tools become simpler to build and less likely to silently give you the wrong number. For anyone who waits days for an analyst to pull a report, the promise is faster, more trustworthy answers straight from the data.