Get AI-Ready With Erik: Creating External Models

Summary

In this video, I delve into creating and using external models to generate embeddings directly within SQL Server 2025, Azure SQL Database, and Azure Managed Instance. While there’s minimal coverage in the full course “Get AI Ready with Erik,” which you can access using a special coupon code found in the video description, I provide deeper insights on how to integrate Python scripts for more efficient embedding generation outside of SQL Server. However, leveraging native SQL Server capabilities is incredibly convenient, especially when setting up your environment properly. I cover essential steps like enabling the External REST Endpoint and creating an external model using OLAMA or another provider. By walking through these processes, you can generate embeddings directly in T-SQL, which integrates seamlessly with transactions and triggers, making deployment simpler and more integrated than external methods.

Chapters

*00:00:00* – Introduction
*00:01:27* – External Models in SQL Server 2025
*00:03:06* – Benefits of Internal AI Embeddings
*00:04:22* – SP Configure Setting Overview
*00:05:37* – Creating an External Model
*00:08:17* – Generating Embeddings with T-SQL

Full Transcript

Erik Darling here with Darling Data, and we are going to talk about creating and using external models to generate embeddings in SQL Server 2025, Azure SQL Database, and Azure Managed Instance also. But there’s a good reason for that. And it is something that I am glad we are able to do, but it takes a little bit more work than you might be into. So, up until now, at least in these videos, there’s been a minimal coverage in the full course, right? The full course, Get AI Ready with Erik, which you can use this wonderful coupon code for, and this coupon code exists down in the video description. There’s a bit more coverage on, you know, including, including like Python scripts and, you know, all the other stuff you need in order to generate embeddings externally and then get them into SQL Server, which is, which is, to be honest, is much more efficient, right? Because there’s all sorts of things you can do external to SQL Server that are a lot faster than the things you can do inside of SQL Server. However, doing things inside of SQL Server has many conveniences, so long as you are able to prep your environment to accommodate such conveniences. So, SQL Server 2025 can generate embeddings natively. We have two new functions, like, well, you know, let’s create external model, which allows us to register a provider with which we can call these things. And we have the AI generate embeddings function, which allows us to generate vectors natively in T-SQL.

Cool. Wait, both of those things sound just dandy, don’t they? So, there are some benefits here. Like, you don’t have to do any Pythoning, which, you know, honestly, if you’re doing this in SQL Server, like, you probably just have Cloud Code write you some Python to do the other stuff, right? Like, you don’t actually have to know Python. You know, you might, you’re going to have to, like, you know, spend some time making sure the robots get stuff right, but, you know, like, the robots are good at stuff like Python, C Sharp, things like that, because, like, those languages are a lot easier for robots to deal with. SQL, they’re terrible with. There’s a lot more, or rather, there’s a lot more debugging with SQL than there is with other stuff. And they can fix the other stuff. Like, if you’re like, hey, robot, run this script, and, like, the script has, like, hits an error, they’d be like, oh, well, I can fix that.

With SQL, you’re like, hit an error, and they’re like, dog. You’re on your own. But you can do some stuff with AI generating beddings, right? Like, works in triggers, works in store procedures. It’s integrated with transactions. And the deployment is a bit simpler. In the full course, I teach you how to do all this. There are some external dependencies that you will need to do this stuff. I think, like, the biggest thing is that OLAMA is only, like, set up to do HTTP requests, not HTTPS. And this stuff requires HTTPS. So you need another program to sort of sit between, like, SQL servers calls in OLAMA to, like, get past the HTTPS requirement there.

In the full course, I teach you how to do all that. In this one, in these short videos, you don’t get the full thing because I still need people to buy the course so I can support the banality of my existence. So we’ve got the new SP configure setting called External REST Endpoint. And this will tell us whether it is enabled or not. We have, of course, enabled this because we’ve done this all before. If you need to change that, you can use, well, this command here. I’m not going to type it for you, but you can do it.

And then you will also need to create an external model. I am not going to drop and recreate mine because I like mine just fine. But this is generally what it will look like. If you were doing this just locally, it’ll look like this. If you’re using, like, OpenAI, there’s other stuff to it. But, you know, I don’t want, like, stop giving Sam Altman money.

Like, my sometimes friend Andy Pablo was posting pictures on Twitter about, like, going to the OpenAI offices. And for a company that lost $9 billion or something last year, they, they, like, if, I could lose $9 billion, too. Like, they bet on that. But don’t use OpenAI. Stop giving that man money.

But this is generally what it will look like if we wanted to do something locally. For the purpose of this course, I’m using this model. So that’s like, you don’t, you won’t use this. You’ll do something else. I’m using OLAMA to do all the processing. And I have a little local webby thing set up to do all this for me.

All right. So this is the, this is the location of the, where I send stuff to get a response back that’s a vector form, that’s a vector data type or an embedding, rather. You can see what models you have registered, right? If you run a fancy query like this, right, and you will see that I, me, I already did all this stuff. So I got all this stuff in there. Man, I did that a while ago, huh? I guess so.

But the end result, once you have done all of this stuff, is that you can generate embeddings in SQL Server. Now, if we go over here, and you can just ignore that, because I had to do something else stupid. We’ve got some other stuff running over here, right? And if we run this query, I got to select the whole thing, I suppose.

If we run this, and we go look at these, well, we should, you know, normally you see a little bit more action in here. You know, that was, that was kind of boring. I apologize. But we get back a string, right? So we said, hey, SQL Server, I want you to generate some AI embeddings for me.

And I want to know what the embedding for how do I optimize SQL Server performance would look like. And I want you to use my, the model that I set up with create external model. That’s this thing up here. And now I get back this, right?

And this is apparently the secret to the universe, is a long list of floats. Look at all these magnificent floaty beasts, these creatures of just fanciful existence. We get back 1024, well, yeah, because they don’t start on line one up here, right?

They start on line two. So because the model that I chose, like, what I tell it to do is generate a 1024 vector embedding. So that’s what I get back.

That’s why we have a, it goes down to 1025 here, but the actual numbers up here start on line two. So that’s what we get back. And this is what we can use to put into our column, columns in SQL Server where we have embeddings, right? And I define those as vector, right now, float32, because that’s what’s generally available.

Float16 is still a preview feature. But I have my vector column set up as float32 with 1024 dimension capacity, right? Or, like, exactitude or something, because I can’t put in less and I can’t put in more.

So this is the string that I get back, and it’s all set up to do the thing that I want it to do. And from there, we can do all sorts of other fun stuff. But additional fun stuff is going to have to wait until tomorrow, right?

So this is a basic overview of how we generate an embedding directly with T-SQL. Thank you for watching. I hope you enjoyed yourselves.

I hope you learned something. And I will see you in tomorrow’s video where we will continue on this path. For those of you who are getting sick of this path, I don’t have that many more videos on the subject. left to release, so your time is almost done.

All right. Thank you for watching.

Going Further

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. Blog readers get 25% off the Everything Bundle — over 100 hours of performance tuning content. Need hands-on help? I offer consulting engagements from targeted investigations to ongoing retainers. Want a quick sanity check before committing to a full engagement? Schedule a call — no commitment required.