How to Set Up Databricks AI/BI Genie

Databricks AI/BI Genie turns a natural language question into a SQL query, runs it against your data warehouse, and returns a readable answer in seconds. Setting it up takes roughly two hours for a single data domain: registering tables in Unity Catalog, provisioning a SQL warehouse, creating a Genie space, and building the knowledge store that keeps responses accurate. Once live, business users can query data without touching SQL or filing a ticket with the data team.

What Genie Does and Does Not Do

Genie is a conversational interface built into the Databricks workspace. It generates SQL based on user questions, executes that SQL against a connected warehouse, and returns results as text summaries, tables, or charts. It is not a general-purpose AI assistant — it only answers questions about the specific tables you add to the Genie space.

This constraint is a feature, not a limitation. Because Genie is scoped to a defined dataset, responses are governed, traceable, and link back to the SQL that produced them. Every answer can be audited by reviewing the generated query.

Genie is available on all Databricks workspaces on AWS, Azure, and Google Cloud. It requires Unity Catalog for data governance and a Pro or Serverless SQL warehouse for query execution. According to Databricks' 2026 release notes, Genie space management APIs are now generally available, making it easier to provision and configure spaces programmatically at scale.

Step 1: Register Your Data in Unity Catalog

Genie only queries tables and views registered in Unity Catalog. If your tables are already there, skip to Step 2.

Unity Catalog is the metadata and governance layer that sits on top of all Databricks data. Tables registered there carry column names, types, and descriptions that Genie draws on when generating queries. Without Unity Catalog registration, a table cannot be added to a Genie space.

To register a table, open the Data section in the Databricks sidebar, select your target catalog and schema, then use the Create Table wizard or run a CREATE TABLE statement pointing to your storage location in S3, ADLS, or GCS. Managed tables, external tables, views, metric views, and materialized views are all supported.

Start with 5 to 10 tables from a single business domain — sales, operations, or customer support — with fewer than 50 columns each. Databricks recommends keeping the initial scope narrow. Genie accuracy drops when the semantic space is too broad and tables from unrelated domains are mixed together.

Step 2: Provision a Pro or Serverless SQL Warehouse

Genie requires a Pro or Serverless SQL warehouse to execute queries. Starter warehouses are not supported.

Navigate to SQL Warehouses in the Databricks sidebar and click Create SQL Warehouse. Select Pro or Serverless as the type, then choose a cluster size. A Small warehouse handles most analyst workloads with response times under 10 seconds for typical business queries. Enable Auto-stop to prevent idle compute costs during off-hours.

Serverless is the practical choice for most teams. Queries start immediately with no warm-up time, and there is no cluster configuration to manage. Pro warehouses make sense if your organization has compliance requirements around compute isolation or needs predictable performance SLAs.

Note the warehouse name — you will need it when creating the Genie space.

Step 3: Create the Genie Space

A Genie space is the container that connects your Unity Catalog tables, your SQL warehouse, and the natural language interface.

Click Genie in the left sidebar, then click New in the upper-right corner. Select the tables or views to include — up to 30 per space — then choose the SQL warehouse from Step 2. Name the space to reflect its domain, such as "Operations Analytics" or "Sales Q&A," then click Create.

After creation, Genie is immediately functional. Type a question and it will generate and run SQL. At this stage, accuracy depends heavily on the quality of table and column metadata already in Unity Catalog. The knowledge store in Step 4 closes the gap between "functional" and "reliable."

Step 4: Build the Knowledge Store

The knowledge store is a collection of curated semantic definitions that tells Genie how your organization talks about data. This is the step most teams underinvest in, and it is the primary reason Genie responses drift from expected results.

Write table and column descriptions in the Genie space editor. Add plain-English definitions to each table and its key columns. For example: arr means "Annual recurring revenue in USD, calculated as monthly recurring revenue multiplied by 12." The more specific the description, the fewer ambiguous-query errors appear in practice.

Add synonyms to map business vocabulary to technical column names. If your team calls the closed_won_amount column "bookings," add that mapping. Genie will then correctly interpret "show me bookings by region last month" without guessing at the column name.

Add example SQL queries covering the most common questions from your team. Include 5 to 10 representative queries. These serve as verified templates that Genie references when generating similar questions. They also constrain the model toward patterns your data actually supports.

Define join relationships explicitly if your Genie space includes multiple tables. Specify how tables relate — for example, orders.customer_id links to customers.id. Without explicit join definitions, multi-table queries are likely to produce incorrect or incomplete results.

Write SQL expressions for business metrics. Define calculated measures — gross margin, churn rate, average order value — as named SQL expressions once, then give each a business name. Genie will use these definitions consistently rather than approximating the calculation from column names on each query.

Step 5: Test Before Sharing

Before giving your team access, run at least 20 test questions that cover the range of queries users will realistically ask.

Start with factual lookups: "How many orders shipped last week?" Move to aggregations: "Which product had the highest return rate in Q1?" Then test multi-table queries: "What is the average deal size for customers who signed up through a partner referral?"

For each response, click Show SQL to review the generated query. If the logic is wrong, the fix is almost always in the knowledge store: a more precise column description, a missing synonym, or an added example query. Do not share a Genie space until test questions return correct results consistently.

Step 6: Share the Space and Assign Permissions

Once testing is complete, click Share in the upper-right of the space editor. Add team members or Databricks workspace groups and assign access levels.

Can view allows users to ask questions, browse results, and export answers. They cannot modify the space configuration, add tables, or edit the knowledge store. Can edit gives data team members the ability to update descriptions, synonyms, example queries, and join definitions.

Keep edit access limited to the two or three data team members responsible for maintaining the space. Open edit access leads to inconsistent knowledge store changes that degrade response quality over time.

Maintaining Accuracy Over Time

Genie spaces require ongoing upkeep. When columns are renamed, new tables are added, or business logic changes, the knowledge store must be updated to match.

Databricks introduced user feedback buttons in early 2026, letting users mark each Genie response as helpful or not. Reviewing flagged responses weekly is the most efficient way to catch knowledge store gaps before they affect the wider team. Assign one data team member as the owner of each Genie space and include space reviews in your regular data governance process.

For teams that need natural language data queries but do not have a Databricks environment, VSLZ supports plain-English questions about uploaded data files with no warehouse or Unity Catalog setup required.

Summary

Setting up Databricks AI/BI Genie takes roughly two hours for a focused data domain. The technical prerequisites are Unity Catalog registration, a Pro or Serverless SQL warehouse, and a configured Genie space. The work that determines response quality is the knowledge store: column descriptions, business term synonyms, example queries, join definitions, and metric expressions. Teams that invest fully in the knowledge store report a measurable drop in ad hoc requests to the data team within the first month of rollout.