Asking a climate database what the weather was like over the North Sea last January should not require a PhD in geospatial programming — but until now, it effectively did.
Researchers have published STRATOS, a Text-to-SQL framework built specifically for the meteorological domain. The system targets data from Copernicus, the EU's Earth observation program, which generates petabytes of climate and observation data in specialist formats like NetCDF and GRIB that general-purpose query tools cannot handle. STRATOS inserts a resolution step before any SQL is generated: a Spatio-Temporal Resolution Agent maps vague natural-language concepts — "the coast of Portugal," "last winter" — to a localized ontology and cross-references external knowledge bases to pin down exact coordinates and time windows. A separate complexity-aware query rewriter then reformulates expensive spatial predicates, cutting execution times from hours to seconds. The team also released an evaluation workload of 7,520 expert-designed query pairs to benchmark symbolic-to-numeric translation, an area prior Text-to-SQL benchmarks largely ignored.
Most Text-to-SQL research has chased general-purpose benchmarks like Spider or Bird, which treat databases as tidy tables of numbers and strings. Climate data breaks those assumptions entirely: a bounding box is not a city name, and "El Nino conditions" is not a column value. STRATOS is a rare attempt to engineer around a domain-specific gap rather than paper over it with a bigger model. If the query rewriting holds up at scale, it could open Copernicus archives to policy analysts and journalists who currently depend on specialists to extract anything useful.
The benchmark release matters as much as the framework — without a shared evaluation standard, every climate-query paper has been grading its own homework.