Google’s AI turns plain English into SQL with record accuracy
Google has quietly pushed the boundaries of how humans interact with databases. Its latest research model, Gemini-SQL2, converts everyday English sentences into working SQL queries with 80.04 % accuracy on the BIRD benchmark—nearly ten points clear of the nearest competitor. Built atop the recently released Gemini 3.1 Pro, the system promises to streamline data retrieval for analysts and end-users alike.
Smarter text-to-SQL for everyone
Current tools often stumble when queries mix slang, typos, or domain-specific jargon. Google says Gemini-SQL2 handles such cases better by leveraging deeper context from the underlying language model. Researchers tested the model against public benchmarks as well as proprietary datasets, noting fewer parse errors and more reliable joins across tables. If the technique scales, it could let non-technical teams ask ad-hoc questions without rewriting queries or waiting on IT.
What this means for cloud analytics
The company hints that Gemini-SQL2 will appear first in Google Cloud services, potentially powering features in Looker, BigQuery and Vertex AI. Early users could see faster dashboard updates and more conversational data exploration. Competitors may need to revisit their own text-to-SQL pipelines, as the 10-plus-point gap suggests a meaningful leap rather than incremental improvement.
Source: The Decoder. AI-assisted editorial synthesis — TechnoExpress.