This is the first of several Fabric first look test drives. In this post, I will focus on Copilot for Power BI, with Data Science and Data Engineering, and Data Factory test drives to follow in subsequent articles.
But before we jump with the first look test, a bit of background, then some pre-requisites, some housekeeping, and then the test drive.
Copilot in Fabric has several features:
Copilot for Data Science and Data Engineering:
Intelligent code completion,
Automates routine tasks, and
Copilot for Data Factory:
Streamlines data wranglers’ workflows, and
Intelligent code completion.
Copilot for Power BI:
Create reports automatically by selecting the topic,
Generate a summary for the report page, and
Generate synonyms for better Q&A.
Before you can use Copilot capabilities in Fabric, there are a few pre-requisites:
You need at least an F64 Fabric, or a P1 Premium capacity.
Note that an F64 capacity equates to just under A$14,700 per month. So, if you are changing your capacity size to test Copilot, make sure to scale it back down, and it is always good practice to pause the capacity when not in use.
Your F64 or P1 capacity needs to be in one of the regions listed in the following link. https://learn.microsoft.com/en-us/fabric/admin/region-availability
Before you can use Copilot capabilities in Fabric for Power BI specifically, there are a few more pre-requisites:
Your administrator needs to enable Copilot in Microsoft Fabric. This can be done with a new tenant setting group, “Copilot and Azure OpenAI Service (preview)”, in the admin portal.
It is possible to enable this for a subset in the organisation, for example to dedicated security groups only, but in this test, I elected to enable it for everyone.
The propagation takes approx. 15 minutes, according to various Microsoft sources, but I think it takes longer as it was not until the following day, that I was able to get the Power BI components to work.
Please also see the disclaimer pertaining to these two settings: https://learn.microsoft.com/en-us/fabric/admin/service-admin-portal-copilot
For this test you need a dataset (recently renamed semantic model) already set up in your fabric workspace.
Important points
Quality of your model design
Data-, semantic-, business- model design has always been extreme important, and can often make or break the success of a data analytics solution. Much has been written over the past few decades about data modelling principles, including dimensional modelling, star schemas, snowflake schemas, business process focussed design, and so on. This is now even more important than before. If you have any hope in Generative AI being able to interpret your data model correctly, you need to understand the core principles, and apply it to your semantic model design. This includes the way you name your entities, hierarchies, and attributes.
Copilot in Fabric is in Preview (as at Jan 2024)
Copilot in Fabric is still in public preview as at the date of the authoring of this article (Jan 24), so you should fully expect some hiccups and accuracy issues. The value for me lies in the art of the possible when this goes into GA.
As mentioned before, also note the disclaimer on data sovereignty while this is in public preview and not yet available in all regions.
Do not forget ethics, frameworks and governance
Similar, please also familiarise yourself with the Microsoft Responsible AI Standard. https://query.prod.cms.rt.microsoft.com/cms/api/am/binary/RE5cmFl
Finally, for all things Fabric, please take the time to invest in the appropriate procedures and frameworks. Something our team can assist you with.
Now, on to the good stuff.
Copilot for Power BI – a first look:
Create reports automatically by selecting the topic - use AI to create reports
Here I will use AI to create reports, by first asking for suggestions as a starting point, and then by authoring the full prompt of what I want. No coding, no visual drag and drop and design. Copilot does it all.
Navigate to your Fabric workspace.
Make sure you use the correct capacity. If you are in doubt as to which capacity applies, select the ellipses, then workspace settings then premium, and you should see either Premium capacity or Fabric capacity enabled.
Now select your Semantic Model that you want Copilot to create the report from. Microsoft claims that you can now “create reports automatically by selecting the topic”. “Topic’, in my view, implies the subject matter of your semantic model. As stated in the important points I made earlier, a great semantic model will have been designed around a very specific business process, domain, unit, etc. Examples of this could be General Ledger, Accounts Payable, Accounts Receivable, Staff Onboarding, Bank Account Application, Mortgage Origination, CO2 Emissions across Building Assets, Works Management, and so on.
The semantic model I will use today is a sales data model I created a while back, including sales transactions, some extended attributes, product hierarchies and location (cities). Here is the model visually represented (as a matter of context) - a pretty simple example.
I select this Sales semantic model, then “Explore this data”, then “Create a blank report”. Note that “Auto-create a report” is also a neat feature but it is not a Copilot and AI feature but rather a vanilla native Fabric Feature that simply aims to represent the data across the semantic model visually (this feature is not discussed in this post).
In the blank report pane, you should now see ‘Copilot’.
Selecting Copilot will firstly show two important Preview disclaimers.
If these are acceptable, continue with “Get started”. Copilot will give you some starter question examples, for example: “Suggest content for this report”, “What is in my dataset?”, or others. And “Create a page that shows…” as the opening statement to building something.
Let’s start with inspiration first by selecting “Suggest content for this report”. Copilot now crafts some compelling suggestions based on what it has learnt of the semantic model (both from its design and its content). The suggestions Copilot made are:
-Product Performance by Category.
-Promotion Effectiveness by Type.
-Store Comparison by Size and Type.
-Sales Trends and Seasonality.
Let’s go with Sales Trends and Seasonality. I can edit this, or, as in this example, simply hit Create. This delivered me a fully interactive report.
I can of course also instruct it with appropriate prompts to create something from scratch. In this example I asked, “Create a page to compare the sales, revenue, and target achievement of stores based on their store IDs.” Here is the result, with each page build adds a new tab to my report.
Generate a summary of the report page.
I will now use AI to craft a narrative from a report page for me.
In an existing report in edit mode, I select the Narrative visual.
I then select Copilot. This now presents some cool options. For example:
-Give and Exec summary,
-Answer likely questions from leadership, and
-Create a bulleted list of insights.
I select “Answer likely questions from leadership”.
Here is the narrative:
Generate synonyms for better Q&A
In the important points section, I described why the quality of your model is now more important than ever, and part of this is the way in which you name your entities, attributes, hierarchies, etc. One way to ensure a proper naming convention is by using synonyms. This will help Q&A work much better and accurately.
It must be noted that the Q&A visual and its natural language processing capabilities aren't reliant on generative AI. However, you can use Copilot for Power BI to quickly improve the Q&A visual's ability to understand user questions.
Adding synonyms for every data entity in your model can be time-consuming, even if they're common synonyms for those names. Let’s see what Copilot can do:
To use this feature, you need to enable it within your dataset.
Now back in your report, select the Q&A visual. You might see a banner at the top of the visual or menu prompting you to improve your Q&A visual by getting or adding synonyms.
If you select Add synonyms, then you will be prompted to review those synonyms.
Conclusion:
I like what I see. If I trust the semantic model, then it seems as if Copilot will do a decent job at re[ort creation, explanations and improving the quality. However, I have not assessed the accuracy of the report, or narratives. And this is simply due to the fact that this is a preview feature and a disclaimer clearly states that there may be accuracy issues.
I do however think that of Microsoft get’s this right, and in combination with the two other Fabric Copilots (i.e. Data Science and Data Engineering, and Data Factory), we may just well start to see some practical areas where AI could assist in the speed of data product delivery.
I will shortly extend these test drives to the other Copilots.
References
In addition to this already mentioned in the article:
Overview of Copilot in Fabric – Overview of Copilot in Fabric (preview) - Microsoft Fabric | Microsoft Learn
Responsible use of Copilot – RE5cmFl (microsoft.com)
Comments