My first comment to the "create a measure description with Copilot" feature just released in #powerbi. PREAMBLE I haven't seen any LLM writing decent DAX so far - they're barely good for a demo, but they are not helpful in the real world. However, I was impressed by ChatGPT-4's ability to explain existing DAX code. Thus, I'm not impressed by the good (but not excellent) results achieved by Copilot. MY NOTES - It works well when measures, tables, and columns have meaningful names: in these conditions, it can extrapolate the "meaning" of the measure and not just an English translation of the DAX code. - When names are not so clear and/or you use acronyms, it is less good. E.g. "Previous Month" is easier to understand than "PM". However, it understands MOM (Month Over Month) and YOY (Year Over Year). - It seems not to inspect the content of the referenced measures. I'm not 100% sure, but so far, I would say that it understands the behavior of a referenced measure just by its name, not by its internal DAX formula. - When you use USERELATIONSHIP, it never understands the meaning of what is happening. Writing in the description that it activates a relationship is not helpful to the user who does not see the diagram view. CONCLUSIONS If you use it to describe the measures you know, it saves you time. A lot of time. I would like to control the style: what's the point of starting each sentence with "The 'xyz' measure calculates..."? Just use "calculate". A dictionary with custom acronyms would be helpful. Customizing the prompt with additional instructions would help, too. It would also be helpful to create the description for multiple measures at once, editing them in a table (in Excel, why not), and then apply them all at once. For consistency, you might want to switch back and forth between different measures. The entire point of this feature is to increase productivity - reducing the needed clicks and actions is part of that journey. https://lnkd.in/dytB4Cd4
Hi Marco Russo, thanks for sharing your opinion on the matter. What is your take on the new hot topic: Visual calculations? Was it really developed the way you initially proposed?https://www.sqlbi.com/articles/a-proposal-for-visual-calculations-in-dax/
Mike Honey
2mo
Thanks Marco, very useful insight, especially for us plebs without access to a P1. I was also expecting it to inspect the functions in the measure, which seemed to be what the screenshots in the blog post were implying (translating DISTINCTCOUNT as "the distinct count") ?
Josue A. Bogran
2mo
Weston Hutchins, thought you might find some parallels with some of the work your team is doing that might be helpful.
Lucio Benussi
2mo
. Visto che sei italiano, magari, qualche volta potresti ANCHE scrivere qualcosa in italiano. Vero? .
Thanks for sharing
Matthew Cymbalak
2mo
Am not surprised given how LLMs work and the nature of DAX. And we must wonder what was the state of the training data!
Moisés Gonga
2mo
Marco Russo... Most of the time, Chat GPT is great for explaining an idea or a piece of code. However, in many cases, the majority of the DAX code doesn't work properly.
Love this
Power BI Architect | Microsoft Fabric | Azure | Databricks | DAX | Cognos Analytics | SQL | 14 Yrs of Exp.
2moFor some reason, this feature is not getting enabled on my pbi desktop even though, I have enabled this option at tenant settings.