Google launches QueryData

Asking a database a question in plain English sounds simple, but the system still has to write exact Structured Query Language, or SQL, without pulling the wrong rows. Google Cloud said on April 10 that its new QueryData service is in preview for that job across AlloyDB, Cloud SQL, and Spanner. (cloud.google.com) QueryData turns natural-language prompts into database queries and is aimed at software agents that need to fetch live business data before taking action. Google said it supports AlloyDB, Cloud SQL for MySQL and PostgreSQL, and Spanner for GoogleSQL. (cloud.google.com) The core problem is that a large language model may know SQL syntax but not a company’s table names, business rules, or which values in a column actually matter. Google said QueryData uses “context sets” — collections of templates, facets, and value-search hints — to map a prompt onto a specific database schema. (docs.cloud.google.com) Google’s documentation says those context sets can include authored examples, schema details, and value searches that trigger automatically when the system needs to match a term to data stored in the database. The company also says developers can download and edit the context file when generated SQL misses the mark, then upload a revised version. (docs.cloud.google.com, docs.cloud.google.com) Google is pitching that setup as a way to make database-backed agents reliable enough for production, not just demos. In its launch post, the company said the target is “near-100%” accuracy for natural-language-to-query tasks, building on its performance in the BiRD benchmark for text-to-SQL systems. (cloud.google.com, cloud.google.com) The timing fits a broader shift inside Google Cloud toward agent software that can read data, decide what to do next, and call tools on its own. Google has also been pushing evaluation tools such as Prism for conversational analytics agents, a sign that it expects customers to test these systems before wiring them into real workflows. (cloud.google.com) Google said Hughes Network Systems is already using QueryData in production, giving the launch an early customer reference even though the product itself is still in preview. The company has also published step-by-step guides showing developers how to connect QueryData to applications through Model Context Protocol, or MCP, tooling. (cloud.google.com, docs.cloud.google.com) For customers, the practical pitch is narrow: let an agent ask the database a question, but fence in how that question gets translated and executed. Google’s bet is that better constraints, more database-specific context, and a human-editable feedback loop will make that safe enough to use on production data. (cloud.google.com, docs.cloud.google.com)