Building a Production Ready Genie Space

What it is

A Genie space lets business users ask questions about data in natural language, and Genie answers by writing SQL over the tables you curate. Newer Databricks documentation calls it a Genie Agent; the exam uses Genie space.

Genie's context is the curation you give it: Unity Catalog tables and views, general instructions, sample questions, SQL expressions and trusted assets. A space is production ready when each element does its own job and a benchmark set shows the answers are accurate.

How it works

The curation elements

Genie does not use volumes or whole catalogs. It works from the specific tables and views you add, so keep the set focused on the domain.

Matching the element to the symptom

Sample questions do not teach Genie terminology. They guide users, so they are the wrong fix when the problem is how Genie interprets a word.

Databricks advises keeping text instructions limited and focused, because too many reduce their effectiveness in longer conversations. Where a term has one exact calculation, prefer a SQL expression or example SQL, since structured definitions are more reliable and easier to maintain.

Trusted assets

A trusted asset is a parameterized example SQL query or a Unity Catalog SQL function that you have checked. When Genie answers with one, the answer carries the Trusted label, telling the user a curated query produced it.

CREATE FUNCTION finance.treasury.desk_exposure(desk STRING)
RETURNS TABLE (trade_date DATE, exposure DOUBLE)
COMMENT 'Daily exposure for one desk, by trade date'
RETURN SELECT trade_date, SUM(notional)
       FROM finance.treasury.trades
       WHERE business_unit = desk GROUP BY trade_date;

Use trusted assets for the few questions where a wrong answer is costly. They do not cover open ended questions, so the rest of the space still needs instructions and good metadata.

Validating with benchmarks

A benchmark set is a list of realistic questions, each with a reviewed correct answer. Run it before release and after every curation change, and compare accuracy instead of judging from a few manual chats.

Worked example: a benchmark of 40 questions scores 28 correct (70%). Nine failures are vocabulary misreadings and three involve one quarterly report. You add four lines of general instructions, the rerun scores 37 of 40 (about 93%), and the quarterly report behind the remaining three failures becomes a trusted asset.

Change one element at a time. If you add instructions, examples and new tables together, you cannot tell which change helped or hurt.

How the exam asks about this

• "Genie answers numbers correctly but does not understand the team's jargon." Add general instructions that define the vocabulary.
• "Regulatory figures must come only from a reviewed query, with a visible marker." Register a trusted asset, a parameterized query or Unity Catalog SQL function, so answers show the Trusted label.
• "How should the team confirm the space is ready for production?" Run a benchmark set of reviewed questions and track accuracy across changes.