Google launches QueryData

Google Cloud has launched QueryData in preview, a service meant to turn plain-English questions into database queries without letting artificial intelligence agents improvise unsafe SQL. (cloud.google.com) Databases answer questions through Structured Query Language, or SQL, the command language that tells a system which rows to read, filter, or join. QueryData sits between a user’s request and the database, then generates a query for AlloyDB, Cloud SQL for MySQL and PostgreSQL, or Spanner for GoogleSQL. (cloud.google.com; docs.cloud.google.com) Google said on April 10 that QueryData is designed for “near-100% accuracy,” and tied the product to its performance on the BiRD text-to-SQL benchmark. The company said Hughes Network Systems is already using QueryData in production for support operations. (cloud.google.com; bird-bench.github.io) The core problem is that writing valid SQL is not the same as writing the right SQL for a specific company’s data. Table names can be cryptic, business rules may live outside the schema, and a query that runs successfully can still return the wrong answer. (cloud.google.com) Google’s answer is a “context set,” a package of JSON files that describes how a database is organized and how the application is supposed to ask questions. Developers build those files with Gemini command-line tools, upload them to Google Cloud, and use them to steer QueryData toward approved query patterns. (docs.cloud.google.com; docs.cloud.google.com) That design shifts some work from the model to the developer. Google’s documentation says teams must enable the Data Analytics application programming interface with Gemini, the Gemini for Google Cloud application programming interface, and the Dataplex Universal Catalog application programming interface before wiring QueryData into an app. (docs.cloud.google.com) Security is the other half of the pitch. Google said enterprises need deterministic access controls rather than relying on a model’s judgment, and it pointed to parameterized secure views that pass fixed values such as a user identifier or region separately from the generated query. (cloud.google.com; docs.cloud.google.com) Google has been moving its data products in this direction for more than a year. In April 2025, the company recast BigQuery as an “autonomous data-to-AI platform” and said usage of Gemini code assist in BigQuery had grown 350% over nine months, with a code-generation acceptance rate above 60% across SQL and Python. (cloud.google.com) InfoWorld framed QueryData as a way to put policy controls around a risky interface: artificial intelligence agents touching structured systems that hold pricing, inventory, balances, and transactions. QueryData does not remove that risk, but it shows where Google thinks the next fight is: not getting models to write SQL at all, but getting them to write only the SQL they are allowed to write. (infoworld.com; cloud.google.com)