Are you using ChatGPT in the Data Intelligence space?

100% agree - thanks for the comment! I personally really like the efficiency + human oversight

Interesting find on Trident.AI, just watched it - what we do differently in the blog is adding business context on where the column is located context-wise (what table, schema,…) and I also prefer to (by default) enable the user to change the suggested description. In that way we have human oversight on what goes in the catalog

@arthur.burkhardt what are you up to with the GPT buzz?

I love DeepL @martijn.vanhauwaert!
Didn’t know they had an API, great, I should look into it

I’m even wondering, would DeepL or gpt-3.5-turbo be cheaper to use? Any findings?

Hi @alexandre.tkint.collibra.com ,

DeepL has a free tier that allows up to 500k characters per month. gpt-3.5-turbo costs $0.002/1k tokens (1k tokens = 750 words roughly).
Further the client we implemented the Deeple solution for was already working with Deepl and had some licenses available. So that made the choice easier

We also tested out ChatGPT to propose business term definitions, but so far the results were a little bit disappointing. Especially for organization specific terms. We are looking further into it how we can improve the results by giving a more specific request.

I’ve found some success in getting ChatGPT to produce good definitions. Here are some things to try:

1. Provide it with your definition rules/standards.
2. Provide it with an example or two of a well-formed definition.
3. Provide it with an example or two of a poorly-formed definition - you can adjust these based on the types of errors you see it return.
4. Instruct it to return 3 competing definitions for a user to choose from.
5. Provide it with a reasonable amount of context.
6. Allow the user to provide up to 5 keywords.
7. Allow the user to explain in natural language why the returned text doesn’t hit the mark.

Organization-specific terms can certainly be challenging for a LLM model as it’s not going to have access to the company’s internal materials, but by incorporating some of the ideas above I believe you’ll at least have a good starting definition for many of them.

I like these suggestions @wes.arnquist.solidigm.com!! Especially providing examples of good / bad definitions

Thank you!! Hope this helps for all data citizens here

That sounds like a really good suggestion @vshahinyan.questrade.com, on what kind of assets/relations would you focus on a first trial?

Thanks @alexandre.tkint.collibra.com, I think we can target anything, e.g. ‘Customer report’ can trigger filtering the result only for reports… actually any improvement case that can lighten search output processing for consumers would be great.

How is that going so far? Any progress? I’m very curious!