Helping a student query data with Genie: from table setup to charts

I hadn't touched the Genie tool before, so I gave Databricks' Genie a try.

First, let me explain what it is. You have a table in hand, dozens of columns, thousands of rows. Previously, to get "how many records total last month," you'd have to write SQL—that's a command that makes the database do work, and one wrong comma in the syntax and it won't run. What Genie does is: you ask in Chinese, it translates that Chinese into SQL, runs it, and gives you a result table, sometimes even drawing a chart for you.

I spent an afternoon walking a first-year master's student in our group through it. Below is the process.

1. Enter the workspace. Find the Genie entry in the left sidebar, click New, and create a Genie space. The literal translation of this word is space—just think of it as a query room. Put the same batch of related tables in the same room, don't mix them.

2. Attach data. Drag the tables in under Add tables, then go to the metric definition column and write the Chinese names and definitions of the fields. After writing, a list of fields appears on the right, each field followed by its own description.

3. Ask the dumbest question first. Don't write select count(*), just type "how many records total does this table have" and hit enter. The interface splits into two parts: the SQL it generated on top, the result below. From my testing, the first question basically always passes.

4. Then ask a business question. "How many records per day last month." It generates a SQL with group by, and the result is a table sorted by day.

5. Make a chart. The top right of the result table lets you switch views—switch to chart, pick line or bar yourself, don't use its default.

6. Follow up. In the same dialog box, continue asking "keep only Wednesday's data," and it will modify based on the previous round's SQL, no need to start over.

Going through the process, the pits you're likely to step in are these. First is searching the wrong thing—search for Genie tutorials and half the results are about Google DeepMind's Genie 3.

Genie 3 is a world model that generates an explorable 3D world from a single prompt, outputting 720p at about 24 frames, and it's said to remember objects placed in the scene for about a minute.

That's a completely different thing from Databricks' query Genie. I got sidetracked for ten minutes at first too.

Second pit is definitions. You ask "how many active users are there," and the model doesn't know how your group defines active—does logging in count, or does running a task once count. This must be written into the metric definitions in advance, otherwise it gives you a number, you use it to write a paper, and when a reviewer asks how this number was derived, you can't answer.

Third pit is the data itself. Date formats aren't unified, units are sometimes milliseconds sometimes seconds, missing values are filled with empty strings. It won't discover these problems for you—though it reports errors pretty fast. Running an OCR pipeline gave me the same feeling—none of the upfront cleaning effort can be saved.

Fourth pit is permissions. If the table isn't authorized to you, clicking in just gives an error—nothing to do with the model, go find whoever manages the data.

Compared to writing SQL yourself, the differences are in a few places. Onboarding threshold: writing yourself requires learning syntax, using it just requires Chinese. Definition consistency: writing yourself relies on your own memory, using Genie you write it in the definitions and the whole group shares it. Multi-table joins: writing yourself is controllable, using Genie easily guesses wrong, needs review. Result credibility: writing yourself depends on the person, using Genie depends on how detailed the definitions are. Overall, the advantages are low threshold, fast follow-ups, and definitions have a place to live; the disadvantages are it guesses when definitions aren't fully written, cross-table joins easily go wrong, and results need a manual review pass. Last week I wrote a piece on agents getting things done, don't rush to pay—the idea is the same: first make results reviewable, then talk about automation.

Next step could be connecting the same query room into a dashboard, so the weekly report for group meetings auto-refreshes instead of manually running it each time. I judge that in the next year or two, asking data questions in Chinese will become the default entry point for data platforms, and writing SQL will recede to a review skill. But whether the numbers count still depends on who writes the definitions clearly.