Genie Space Configuration: Instructions, Example SQL and Trusted Assets

1.Before you curate: data, compute and permissions

The exam says "Genie space". The current documentation says "Genie Agent" and notes that Genie Agents were formerly known as Genie Spaces. Both names refer to the same object. Databricks describes it as a natural-language chat interface. Data analysts configure it with Unity Catalog data, example SQL queries, instructions and trusted assets, and business users then question it through the chat UI.

Before you add any instructions, the space has hard requirements. All of its data must be registered to Unity Catalog. That includes managed, external and foreign tables, views, metric views and materialized views. A space can hold up to 50 tables, views or metric views. It also needs a pro or serverless SQL warehouse, and Databricks recommends serverless for the best performance.

Credentials work in two ways, and exam questions often test the difference. When you save a warehouse, your compute credentials are embedded in the space so every user can reach that warehouse. Those embedded credentials cover the warehouse only. Data access is always checked against each end user's own Unity Catalog permissions. If another author later changes the warehouse, their credentials are embedded instead. Genie also relies on partner-powered AI features, which must be enabled at both the account and workspace levels.

Sources1

2.Scope the domain: one audience, few tables, good metadata

Selecting assets for a business domain starts with deciding who the space serves. The documentation says to define the audience and purpose first, because that choice determines which data, instructions and test questions belong in the space. Treat 50 tables as a ceiling, not a target. The guidance is to start as small as possible and aim for five or fewer tables. Limit the columns too, and hide any column that is confusing or unimportant. If the topic seems to need many tables, pre-join them into views or metric views. Metric views work especially well because they pre-define metrics, dimensions and aggregations.

Genie writes SQL from Unity Catalog column names and descriptions, so the metadata on those tables is your first set of instructions. Descriptions should be precise and contextual. If you use AI-generated descriptions, check them for accuracy and keep them only if they match what you would have written yourself. You can also add column synonyms and custom descriptions inside the space. That metadata applies only to the space and doesn't overwrite what is stored in Unity Catalog.

Choosing data is therefore a decision about attachment. Anything Genie needs must be listed under the space's Sources, and attaching a source still leaves access governed by each user's Unity Catalog permissions. If you skip this decision and ask Genie Code to build the space from a description of your domain, it identifies relevant data and confirms the sources with you before building.

Sources2

3.Pick the right home for each piece of business knowledge

With the data scoped, the next decision is where each business rule should live. The documentation ranks the options. Use SQL expressions in the knowledge store for frequently used metrics, filters and dimensions, such as revenue or gross margin. Use complete example SQL queries for hard-to-interpret, multi-part or complex questions. Use SQL functions registered in Unity Catalog for logic that a static or parameterized query cannot capture. Use text instructions only as a last resort, for guidance that genuinely needs natural language, and never for metrics, filters or query patterns that SQL can express.

The ranking matters because instructions compete with each other. Genie performs best with a limited, focused set of instructions, and too many can reduce effectiveness, especially in longer conversations. Each space also has two separate budgets, and the exam may ask which asset counts against which.

The snippet limit has explicit exclusions. Text instructions, example SQL queries, SQL functions, column descriptions and prompt matching settings don't count toward the 200 knowledge store snippets. Prompt matching is added automatically when you attach tables. It matches user wording to actual column values and corrects spelling, and you can choose which columns it is enabled on.

Sources23

4.Example SQL queries and trusted assets

Databricks recommends example SQL queries as the main way to give a space instructions. Each one pairs a question with the SQL that answers it. Use the most typical way a user would phrase the question as its title, because the title is what Genie matches incoming prompts against. When a question matches, Genie can use the example directly. When a question is only similar, Genie borrows the example's structure. Pick examples that show logic unique to your organization, such as which opportunities count as pipeline.

You don't have to write every example yourself. When you attach data, Genie searches for popular workspace queries on those assets that you have at least CAN VIEW on. It offers them in a Review Suggested Queries dialog, where you accept or reject each one, and accepted queries become fully editable. Genie doesn't suggest queries on tables that aren't attached. If you built a dedicated view for the space, queries that ran against the original source tables won't show up. Queries that only perform basic writes are also skipped.

Trusted assets are the subset that gives users a verified answer. There are two kinds: parameterized example SQL queries and SQL functions registered in Unity Catalog. A parameter is added with a colon followed by a name (such as :parameter_name). It has a data type (String, Date, Date and Time, Decimal or Integer) and a comment that helps Genie pick the right value. Choose the type carefully, because a value that doesn't match the type can lead to inaccurate results. SQL functions handle logic too complex for a query. Genie cannot view or modify the SQL inside them, and users need EXECUTE on any function used as a trusted asset. You can also add usage guidance to an example to tell Genie when it is particularly relevant.

Sources31

5.Text instructions: global, specific and consistent

After SQL has covered what it can, what remains goes in text instructions. General text instructions apply to every prompt. If a rule only matters for some questions, put it in an example query, a function or table metadata instead. Typical general instructions are company facts like a fiscal calendar, plus formatting rules. You can write them as one note or group them by topic, as in this documented example:

- **Company-specific business information**:
  - Fiscal year starts in February, for example fiscal year 26 or FY26 is February 1, 2026 through January 31, 2027
- **Formatting**:
  - Always respond in Spanish
  - If no other specification exists, round all decimals to two places
  - Omit commas in results for any column including "Id" or "id" or "\_id"

Clarification instructions follow a fixed shape. First state the trigger (the topic), then the missing details, then the required action (Genie must ask first), and finally the exact question to ask. Put clarification instructions at the end of the general instructions so Genie gives that behavior priority.

Summaries have their own rules. To customize the natural-language summaries that accompany results, add a section at the end titled "Instructions you must follow when providing summaries", for example to cite the tables and columns used or to state the date range. Only text instructions affect summaries. SQL examples and SQL expressions don't, and summary length and detail level can't be controlled.

Instructions must also agree with each other. Genie is nondeterministic, so conflicting guidance raises the risk of bad answers, and reviewing for inconsistencies is a key setup task. If the text says to round to two decimals, every example query must round to two decimals as well.

Sources32

6.Validate before you share: be the first user

A space is ready for production only after you have tested it. Databricks recommends adding example SQL, instructions and knowledge store configuration before sharing, and making yourself the space's first user. Ask realistic questions, open Show code, and read the generated SQL. If Genie gets it wrong, anyone with CAN EDIT or higher can fix the SQL, run it, and click Add as instruction to save the question and SQL as a new example query. Genie Code can also find the cause of a wrong answer and propose context changes for you to accept or reject.

No. Benchmarks are for evaluation only. Genie doesn't use benchmark questions or their SQL to improve its context, so a question you want Genie to learn from must be saved as an instruction (an example query). Three kinds of "sample question" are easy to mix up. Common questions are optional prompts shown on the chat landing page, and author-defined ones appear before auto-generated ones. Example query titles are what Genie matches prompts against. Benchmark questions only measure accuracy.

After launch, validation continues. User feedback (Yes, Fix it, Request review) doesn't change Genie's behavior by itself. It points you to what to fix. Users with CAN MANAGE review flagged responses and use the Monitor tab to filter questions by time, rating, user or status. They can also run Analyze Agent Usage to surface recurring issues. Treat the space as an ongoing collaboration that builds up knowledge over time, not a one-time deployment.

Sources2145

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Text instructions are the primary way to teach a Genie space your metrics and filters.Why is that wrong?

   Text instructions are a last resort. Metrics and filters belong in SQL expressions, and complex questions in example SQL queries.

   Covered in Pick the right home for each piece of business knowledge

2. 2.Example SQL queries and text instructions use up the 200-snippet knowledge store limit.Why is that wrong?

   Example queries, SQL functions and the general instructions block count toward the separate 100-instruction limit. Only table descriptions, join relationships and SQL expressions count toward the 200 snippets.

   Covered in Pick the right home for each piece of business knowledge

3. 3.Naming a table in the instructions is enough for Genie to query it.Why is that wrong?

   Genie queries only sources attached to the space, whatever the instructions or metadata mention.

   Covered in Scope the domain: one audience, few tables, good metadata

4. 4.Adding benchmark questions with correct SQL teaches Genie to answer them.Why is that wrong?

   Benchmarks only measure accuracy. To teach Genie, save the question and SQL as an example query.

   Covered in Validate before you share: be the first user

5. 5.Any example SQL query Genie learns from produces a verified answer.Why is that wrong?

   Verified answers come only from trusted assets: parameterized example queries whose exact text is used in chat mode, and Unity Catalog SQL functions.

   Covered in Example SQL queries and trusted assets