QueryData launches in preview today. It is a tool for translating natural language into database queries with near-100% accuracy. With QueryData, you can build agentic experiences across AlloyDB, Cloud SQL (for MySQL and PostgreSQL), and Spanner (for GoogleSQL). It builds upon Google Cloud’s #1 spot in the BiRD benchmark, one of the world’s most competitive benchmarks for natural-language-to-SQL – as well as upon Gemini-assisted context engineering.
Developers are already seeing the benefits from QueryData, including Hughes Network Systems, a leader in telecommunications, that deployed QueryData in production. “We have transformed user support operations with Google Cloud’s data agents. At the heart of our solution is QueryData, enabling near-100% accuracy in production. We are excited about the future of agentic systems!” – Amarender Singh Sardar, Director of AI, Hughes Network Systems
The opportunity for agentic systems: from intent to action
Agentic systems are evolving from human-advisory roles into active decision-makers. To execute business actions accurately, agents require precise information from operational databases (such as pricing, inventory, or transaction records).
With requests expressed in natural language, bridging the gap between conversational input and database records is essential. High-quality natural language-to-query capability is a critical requirement for enabling agents to take actions.
The developer’s dilemma: why natural language for agents with databases is hard
Hurdles for agents querying enterprise data are threefold: accuracy, security and ease of use. QueryData addresses all three of them:
-
Accuracy – Inaccurate answers carry a risk of poor business decisions, disappointed end-users or financial losses. In many industries, translating text into SQL with 90% accuracy is simply insufficient for taking action.
-
Security – how to make sure that each person (or agent) only queries the data they are allowed to see? Enterprises need auditable, deterministic access controls. Relying on the LLM’s judgement (aka “probabilistic” access controls) falls short of that. Even a low risk of security breaches means disproportionately high losses
-
Ease of use – Achieving high accuracy requires developers to provide extensive contextual information about their data. This can be a laborious task. Another example of developer friction is integration and maintenance of agentic tools
Understanding the accuracy gap
LLMs are really good at writing query code. However, to write accurate queries for a given database – it takes more than coding skills, and more than just parsing the schema:
-
Schemas can be unclear – developers often use shorthands or abbreviated names. For example: what does a column named “product” mean? A product category? A particular model…? It gets even worse with column names like “prod” or simply “p”
-
Values can be ambiguous – let’s take a column named “order return status”… where values are expressed as integers: “1”, “2” and “3”. Which of these represents “returned” or “return initiated”?
-
Schemas cover data structure, but not the business logic – Your business may define “monthly active users” as those who have posted at least once, not just logged in (but database may lack this nuance).
-
Underspecified queries – Natural language questions can be ambiguous, like “latest sales”.
How QueryData solves for near-100% accuracy
QueryData leverages the Gemini LLM, as well as context which describes your unique database.
Database context, which is essentially the code fueling QueryData, is a set of descriptions and instructions including:
-
Schema ontology – information about the meaning of the data. Descriptions of columns, tables and values. It helps QueryData overcome ambiguity by figuring out what data is needed to answer the question
-
Query blueprints – guidelines and explicit instructions for how to write database queries to answer specific types of questions. Templates and facets specify the exact SQL to write for a given type of question.
As a last resort, QueryData will detect when a clarifying question needs to be asked.
Deterministic security for your queries
Agentic applications require deterministic, auditable security. Developers can use Parameterized Secure Views (PSVs) to define agent access via fixed parameters, like user ID or region. By passing these security-critical parameters separately from queries, the application ensures agents can only access the authorized data. This prevents agents from querying restricted information, even if they attempt to do so.
Support for PSVs is available today in AlloyDB, and coming soon to Cloud SQL and Spanner.
Ease of use for quality hill-climbing and tool integration
Integration of QueryData into your agentic workflows is easy. The QueryData API can be used directly or exposed as a Model Context Protocol (MCP) tool via our popular open source MCP Server: MCP Toolbox for Databases. QueryData automatically works across different database dialects – no need for database-specific code, just one API to query them all.
Another area where QueryData makes things easier for developers – is context engineering. It is the process of iteratively evaluating and optimizing context. It is critical to QueryData’s ability to accurately query your database. Developers using QueryData enjoy support from a robust suite of tools:
-
Out-of-the-box context generation – upon configuring QueryData, the Context Engineering Assistant, a dedicated agent in Gemini CLI, will help you create the very first context set for your database.
-
Evals: Developers can use the bundled Evalbench framework to measure accuracy against a set of tests specific to your use case
-
Context optimization: the Context Engineering Assistant reviews eval results, recommends changes and then helps run evals again. Through this iterative process, you can reach near-100% accuracy.
What you can build with QueryData today
Developers are already building with QueryData. Examples include:
-
Customer-facing applications: a real estate search engine, where QueryData translates user prompts into database queries, and then schedules viewing appointments
-
Internal tools: an AI-powered staffing app querying human resources data and then enabling managers to assign workers to shifts
- Multi-agent architectures: a trade compliance workflow where a top level agent asks a sub-agent to verify that an entity has appropriate KYC (“Know Your Customer”) status. The KYC agent queries a database to confirm the customer’s identity.
Next steps
You can have your agent start using QueryData as a tool for near-100% accurate database calls today. For more details, explore our technical documentation:
Check out the “Swiss property search” high-fidelity demo, pictured below (video walkthrough here). Note: This is an independent project (not maintained by Google Cloud) and is for illustrative purposes only: GitHub link
