ChatGPT and Epicor SQL Query Assistance

With all the AI hype at Insights I have started making an effort to use ChatGPT and test its capabilities. Just thought I’d share this conversation with ChatGPT 4o about a fairly simple Epicor SQL query. It’s not 100% accurate but did pretty well. Some field names are incorrect and the joins aren’t complete but it did help speed up the process of pulling this data.

ChatGPT - Epicor Kinetic SQL Query Assistance

It will be interesting to see how much it learns and remembers over the course of continued conversations. You can see that I asked it about learning from my corrections, and it doesn’t apply them to its overall data model, only to my specific ‘instance’, apparently?

When they start training themselves, is when we need to worry.

1 Like

Given that it doesn’t have knowledge of the Epicor database schema it did pretty good.

I imagine it wont be long before some clever cookie figures out a way to train the model with a specific schema and then creating queries will be much easier…Note: a person who understands the output still needs to review the query generated… remember they are the co-pilot and we are the pilot! dont put a co-pilot in the pilots seat.

You sure about that? Watched this today:

Off Topic, but…

Video

https://www.youtube.com/watch?v=OxDxwXE_Tv0

Off Topic post here:

That episode was brutal. It’s unreal how long that kept going on…

1 Like

That’s called federated learning. Modern AI orgs are extremely cautious about using it. That’s because one time Microsoft was not careful, and ended up with Tay, the Nazi chat bot.

It’s more likely to “learn” Epicor-specific stuff if you post about your queries here. They absolutely crawl this site. I’ve seen explicit examples of Copilot and GPT giving answers sourced from posts here.

Having said that, I vaguely recall @timshuwy mentioning that Epicor itself is training or fine tuning a model solely on Epicor-related topics. That is likely to be more useful and accurate than a massive general purpose model trained on the entire internet.

1 Like

Yeah I can see how it would be dangerous to take information from anyone and add it to the primary/global data model. Anyone could feed it with whatever they want.

If you saw the demo on main stage during the keynote at Insights, you saw the person build a dashboard by giving verbal commands, and it automatically built and produced a color coded dashboard. :safe_harbor: Safe Harbor :safe_harbor: DId you know that 99% of what you saw was totally real code (not just pictures or videos)? Think about what it did… it had to create a query/baq… how did it do that? well, we are training the AI with our data model so that it knows best how to build out the dashboards. But we also need to train it in proper business language. If you ask it “what were my SALES for last month”, it needs to know what you mean by sales.
So, what you saw on the stage was very advanced, and still needs some final touches before it it’s release, but it is very close. :safe_harbor:

2 Likes

I seen the US version of this one years ago.

I’ll pop up a post in off topic post soon, there are some real good ones.

Off Topic post here:

Back to regular scheduled programming.

1 Like