Genie space monitoring: track questions, feedback, and fixes

1.The feedback loop starts with one question: Is this correct?

The exam guide says "Genie space", but the current documentation says "Genie Agent" and notes that Genie Agents were formerly known as Genie Spaces. The two names mean the same product. Optimizing a space starts with finding out where its answers go wrong, and the first source of that information is the users. Each response asks the user "Is this correct?" and gives three options. Yes confirms the answer. Fix it flags the answer as incorrect. The user picks a common issue or writes their own explanation, then chooses either *Submit and try again* to regenerate the answer using that feedback, or *Submit* to send the feedback without regenerating. Request review flags the response for manual review and can include an optional comment.

Feedback is a signal for the people who curate the space. It does not retrain anything. Even so, Databricks recommends encouraging users to give feedback through this mechanism. When you bring in business testers, tell them that their job is to help refine the space. Ask them to upvote the final query once Genie gets it right, and to send unresolved questions straight to the space's authors.

Sources12

2.Tracking questions and accuracy on the Monitor tab

Feedback reaches editors through the Monitor tab. It lists every question asked in the space along with Genie's answer, and you can filter the list by time, rating, user, or status. Users with CAN MANAGE permission use it to see what business users are asking and how Genie responded. Click a question to read the full chat thread. The questions Genie handles badly show you exactly where it needs more specific instructions.

Two features summarize activity across the whole space. The Weekly digest section shows weekly message volume, active users, and thumbs up/down feedback. Analyze Agent Usage opens Genie Code, which reads the last six months of user messages, feedback, and issues. It reports common topics, recurring issues, and suggested context improvements, and each finding links back to the conversations it came from.

Sources1

3.Turning a wrong answer into new context

Once you have found a bad answer, open its SQL. Click Show generated code to see the query. The people who build a space usually know the data well enough to spot the mistake, and a small edit often fixes it. If you have CAN EDIT or higher permission on the space, you can edit the generated SQL and run it. Then click Add as instruction, which opens the example-SQL editor pre-filled with the question and the SQL. The response's kebab menu also has Add as benchmark, which saves the question for systematic accuracy testing later.

The other route is Genie Code. Open it from the response, describe the problem and the behavior you want, then review the context changes it proposes and accept the ones you want to keep. The documentation's example is a space using calendar quarters when the business uses a fiscal calendar. You can also ask Genie Code to capture new terms or corrections that users introduced during a conversation.

Sources1

4.Updating instructions and trusted assets from stakeholder input

Stakeholder feedback usually reaches you as a sentence, such as "that's not how we count active customers." You then decide which kind of context will fix it. The best-practices guidance gives a clear order of preference. Use SQL expressions (measures, filters, fields) for common business terms. Use example SQL queries for questions that are complex, multi-part, or hard to interpret. Use text instructions only when neither of those can meet the need. General text instructions apply to every prompt, and too many of them can make Genie less effective, especially in long conversations. Whatever you add must also agree with what is already there. If the text instructions say to round to two decimals, the example queries have to round to two decimals as well.

Trusted assets are parameterized example SQL queries and SQL functions registered in Unity Catalog. When Genie answers using one of them, the answer is marked as verified. Users with at least CAN EDIT permission can add or remove trusted assets. Users of the space need EXECUTE permission on any SQL function used as a trusted asset. Title each example with the way a user would most typically ask the question, and focus on logic that is specific to your organization, as in this example:

-- Return our current total open pipeline by region.
  -- Opportunities are only considered pipelines if they are tagged as such.
  SELECT
    a.region__c AS `Region`,
    sum(o.amount) AS `Open Pipeline`
  FROM
    sales.crm.opportunity o
    JOIN sales.crm.accounts a ON o.accountid = a.id
  WHERE
    o.forecastcategory = 'Pipeline' AND
    o.stagename NOT ILIKE '%closed%'
  GROUP BY ALL;

Sources23

Exam traps

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

1. 1.When users click Fix it or thumbs-down, Genie automatically learns and stops making that mistake.Why is that wrong?

   Feedback does not change the space's behavior by itself. Editors read it on the Monitor tab and turn it into instructions, examples, or metadata.

   Covered in The feedback loop starts with one question: Is this correct?

2. 2.The quickest way to fix a misunderstood metric is to add a sentence defining it in the general text instructions.Why is that wrong?

   Metrics, filters, and query patterns that can be written in SQL belong in SQL expressions or example queries. Text instructions are a last resort, and too many of them can make Genie less effective.

   Covered in Updating instructions and trusted assets from stakeholder input