Random Procrastination

Darren Gosbell

Tag: ChatGPT

Generating Measure descriptions with ChatGPT – part 2

February 17, 2023 / / 13 Comments

In my previous post I had a short 40 line script for Tabular Editor (free | paid) which could generate descriptions for the measures in your tabular model. There were a number of interesting questions in the comments to that post which I thought would make a follow up post worthwhile.

The things I want to cover in the post are

How to run the script using TE2

Below is a very brief tutorial on running a script using Tabular Editor 2

  1. If you launched Tabular Editor from the External Tools ribbon in Power BI desktop you can skip this step, otherwise click on this icon to connect to your data model.
  2. Click on the C# Script tab and paste in the script from my earlier post
  3. There are instructions in my earlier post for requesting your own API key from Open AI you need to paste the key into this line between the quotes (“)
  4. Optional: if you are targeting users that speak a language other than English you can update the question text on line 10 to your desired language (from what I understand ChatGPT understands about 100 different languages so you might need to test if this works in your language)
  5. Click the play button to run the script, this step could take a while depending on the number of measures and whether you hit throttling (more on this below)
  6. Once the script has run you can check some of the descriptions by clicking on a measure in the model explorer. At this point all the descriptions are still local and have not been saved back to your data model.
  7. The text generated by ChatGPT will appear in the Description property for the measure.
  8. If you are happy with the generated descriptions clicking the Save Changes button will save the new descriptions back to your data model.

How to run the script using TE3

The steps for Tabular Editor 3 are almost the same with the main difference being at step 2 because TE3 allows you to have multiple C# scripts open at the same time.

  1. If you launched Tabular Editor from the External Tools ribbon in Power BI desktop you can skip this step, otherwise click on this icon to connect to your data model.
  2. Click on the New C# Script button to create a C# script document and paste in the script from my earlier post
  3. There are instructions in my earlier post for requesting your own API key from Open AI you need to paste the key into this line between the quotes (“)
  4. Optional: if you are targeting users that speak a language other than English you can update the question text on line 10 to your desired language (from what I understand ChatGPT understands about 100 different languages so you might need to test if this works in your language)
  5. Click the play button to run the script, this step could take a while depending on the number of measures and whether you hit throttling (more on this below)
  6. Once the script has run you can check some of the descriptions by clicking on a measure in the model explorer. At this point all the descriptions are still local and have not been saved back to your data model.
  7. The text generated by ChatGPT will appear in the Description property for the measure.
  8. If you are happy with the generated descriptions clicking the Save Changes button will save the new descriptions back to your data model.

Dealing with Rate Limiting

If you have not worked with API calls before you may not have come across the term “rate limiting” before. Basically, it is a mechanism that APIs use to prevent users from monopolizing the resources on a service by limiting the number of calls that can be made within a given timeframe.

In the case of the Open AI APIs they document their rate limits here and at the time of writing they only allow 20 calls per minute for a free account. Once you hit that limit the API will return a 429 error code which is a common code meaning “Too many requests”. There are 2 approaches that you can use to work around this.

  1. You can add pauses in the code to wait until the next minute when you can then make another 20 calls.
  2. You can upgrade to a paid account which has much a higher limit.

If you would like to see an example of an updated script which will skip measures which already have descriptions (so if you’ve manually updated some or if there was an issue part way through running a previous script

Update Tabular Model Descriptions from ChatGPT with rate limiting logic (github.com)

A Final Warning

The Large Language Model (LLM) behind ChatGPT is an amazing piece of technology and it’s only going to get better over time. But I’ve also seen it described as “a B grade intern who hallucinates occasionally”. In it’s simplest form all these AI models do is to break a piece of text into a series of tokens and then predict what tokens are likely to come next based on a corpus of training material.

In my testing I’ve seen a number of examples where the prediction has missed some important detail. And in a couple of cases, I’ve even seen it include a modification to the original measure and then it has described the modified measure, not the original one!

So, while this is definitely a huge time saver, just be aware that it’s not perfect. Before publishing these descriptions into production, I would suggest doing a review of them to ensure that they make sense and include all the important details.

If you want to see all the descriptions at once you can use a tool like DAX Studio to run the following query to list them. 

SELECT [Name], [Description], [Expression] FROM $SYSTEM.TMSCHEMA_MEASURES

Automatically generating measure descriptions for Power BI and Analysis Services with ChatGPT and Tabular Editor

February 15, 2023 / / 36 Comments

Note: Make sure to go on and read part 2 after reading this which has some improvements to the script.

It seems like everyone is playing with ChatGPT at the moment and doing everything from getting it to write code for them to asking it to write ode’s to their favourite visuals. Writing DAX code is interesting from the perspective of someone building a data model. But one thing that has always been hard from the perspective of consumer of a data model is understanding the logic behind those DAX calculations.

As you’ve probably already guessed from the title of this post rather than generating DAX I wanted to see if we could go the other way and get ChatGPT to provide simple descriptions of what a DAX expression was doing.

To test this I did not want to use a basic measure, so I found a moderately complex expression that someone else had actually originally had ChatGPT generate for them and ask it to explain the calculation. The first time I tried this it produced a very long description breaking down each of the function calls. But by revising the way I asked the question I eventually got to a point where it was giving me a nice concise explanation.

When I posted the following question to ChatGPT:

Explain the following calculation in a few sentences in simple business terms without using DAX function names:
CALCULATE (
   SELECTEDVALUE ( 'Suppliers'[CompanyName] ),
   TOPN (
      1,
      SUMMARIZE (
         Products,
         'Suppliers' [CompanyName],
         "@StockValue", SUMX( 'Products'[UnitsInStock] * 'Products'[UnitPrice] ),
      [@stockValue],
       DESC
   )
)

It came back with the following answer which I was pretty happy with:

This calculation is finding the company name of the supplier with the highest total value of stock for a given product. It does this by first multiplying the units in stock by the unit price for each product, then summarizing that total value by supplier. It then finds the top supplier by that total value and returns their company name.

This was great for a single measure, but not all of the consumers of your data model will have access to view the DAX expression. And even if they did it would be a pain to have to keep jumping out to an external website to understand what every measure is doing.

So this made me wonder if there might be some way of injecting these descriptions into the data model. And as it turns out ChatGPT already has a REST API. All you need to do is to sign up for an account at https://platform.openai.com (if you’ve been experimenting with ChatGPT you probably already have an account) and then generate an API key for your account and you can make requests of ChatGPT from another program.

View and Create API keys from the account link in the top right corner

Note: free accounts are limited to 20 calls per minute (see Rate Limits – OpenAI API ). So for large models you would either need to add logic to include a 1 minute delay every 20 measures or upgrade to a paid plan.

From there I setup the following Tabular Editor script which will loop through all of the measures in your data model and update the description with the one generated by ChatGPT. All you need to do to run this is to paste your own API key into the apiKey constant on line 8.

For illustration purposes I’ve also included the original DAX expression, but you don’t need to keep that if you don’t want to.

View this gist on GitHub

Note: the script has been tested in both TE2 (free) and TE3 (paid)
and is available as a gist on github https://gist.github.com/dgosbell/f3253c7ec52efe441b80596ffddea07c

The updated script including logic for dealing with rate limiting is included in my following post here

Before running the script hovering over a measure in my data model only shows the measure name

the default tooltip showing just the measure name

After running the script, you get a much more informative tooltip

the expanded tooltip showing the new auto-populated description

Interestingly the descriptions do seem to change a little bit from one run to the next and in spite of asking it not to use DAX function names in the description sometimes it still sneaks one in. The description below was particularly bad where it just described the SUMX function.

But the advantage of this technique is that you can re-run it for single measures if you like or you could manually edit the description field if the description needed a small improvement.

Update: 16 Feb 2023 – added note about API rate limiting

Update: 17 Feb 2023 – See part-2 here for short tutorials on how to run the script and an updated version which deals with the API rate limiting

© 2024 Random Procrastination

Theme by Anders Noren — Up ↑