Genie Agents Purpose and Setup

What it is

An AI/BI Genie Agent (formerly a Genie space) lets business users ask questions about data in plain language. Genie turns each question into SQL, runs it on a SQL warehouse and returns a table, a chart or a short text answer. Users can ask follow-up questions in the same conversation.

An analyst builds the agent for one business domain, such as sales or support. The analyst chooses the data, writes the context that explains the business, and vets the queries that must always be answered the same way.

How it works

Components of an agent

Genie only queries the data attached to the agent. A table named in an instruction but not attached is never queried.

Curating the data

Keep the agent focused on one domain with a small, clean set of tables or views. Good table and column comments in Unity Catalog help Genie pick the right columns, and the agent can add its own column descriptions and synonyms. Hide columns users never need.

Views and metric views are useful when the raw tables need joins or business logic first. They also let the author expose only the columns that belong to the domain.

Choosing the SQL warehouse

A Genie Agent needs a pro or serverless SQL warehouse, and Databricks recommends serverless for performance. The author needs CAN USE on the warehouse to set it as the default.

Sample questions and instructions

Sample questions are set as common questions in the agent settings. They appear on the chat landing page and show users what the agent can answer, so they should be realistic questions the data supports.

General instructions should be short, specific and global. A rule that applies to only some questions, such as how to calculate churn, belongs in an example SQL query, a SQL function or a column comment instead.

When you create an agent, Genie Code (Databricks Assistant) can suggest table descriptions and example queries. Genie can also suggest popular queries from workspace query history, which the author accepts, edits or rejects.

Trusted assets

Trusted assets are the example SQL queries and SQL functions that the author has vetted. When Genie answers a question with one of them, the user gets a verified answer instead of newly generated SQL.

Parameters make an example query reusable. A parameter is written as a colon and a name, and has a data type such as String, Date or Integer.

-- Example SQL query: revenue for one region
SELECT order_month, SUM(revenue) AS revenue
FROM sales.gold.monthly_revenue
WHERE region = :region
GROUP BY order_month
ORDER BY order_month;

Users need EXECUTE on any SQL function used as a trusted asset. Only users with at least CAN EDIT on the agent can add or remove trusted assets.

How the exam asks about this

• "Executives must get the exact same revenue number every time they ask." Add a vetted, parameterized example SQL query as a trusted asset.
• "Genie misreads 'active customer' across many questions." Add the definition to the general instructions or as a SQL expression.
• "Which compute can a Genie Agent use?" A pro or serverless SQL warehouse, serverless recommended.