Today I had the pleasure of presenting my session on Copilot for SQL at the Data Ceili event in Dublin. Somewhere between 50 and 60 people attended, having fun and learning how this Copilot works, what it can and can’t do, and, very important today, how much it will cost you.
The challenge
While preparing this session, I tested some features in SSMS 22.6, and everything looked good.
Then, SSMS 22.7 dropped! With some major changes. Agent mode, for instance, as well as SQL formatting.
The evening before the session, Erin Stellato posted a blog explaining that, apart from a Copilot instructions file, you can now also bind Copilot in bounds to prevent it from doing things you do not want it to do.
Finally, Microsoft announced a new billing model for GitHub Copilot, which has led some people to run out of tokens in a few days.
All these changes together made presenting the session both exciting and challenging. Mostly because I wanted to show these new features without breaking anything.
I do not mind it when demos fail; I do when I can’t figure out why. I had to pause some testing as I didn’t want to run out of tokens, for instance.
The choices
After the speaker dinner, an idea started to form in the back of my head. The next morning, whilst chatting with others, the idea fully formed. When I went up to the speaker room, I knew what to do. Add action music here.
I decided on showing the new features of SSMS. Also, I’d show the Copilot instructions file. What I would skip was creating the specific Copilot account. Why? Well, it’s very likely that I would mess something up, and I just didn’t have the time to fix it.
The result
In the session, I explained the new features, showed them where possible and explained how they help GitHub Copilot. This way, I worked around potential issues without taking away too much information for the attendees.
I ran the demos (typing live), and we had fun with the results. Not only because there were some errors, but also because of the ever-optimistic tone of the LLM. One reason errors came up was that the prompts were vague. On purpose, to show the importance of writing good, specific prompts.
One of the new skills you need to learn.
I think it all worked out; attendees hopefully learned how this agent can assist them in their work, that there is an option to limit what it can do, and how to provide more information to the agent and the model to improve the results. I’m also still convinced that it’s a great help, but won’t take away your expertise and knowledge.
Oh, and if you’ve read carefully, you saw the mention of Copilot token usage. My level went up from 0.7 per cent to a whole 1.2 per cent.
I was using auto mode, which should choose the best fit for the question asked. The funny thing is, when you ask Copilot which model it is using, it won’t show that. But it is shown in the sidecar chat.
Finally, if you have the time, take some time out and visit Data Ceili in Dublin. It’s a great event with a very relaxed vibe at a lovely location.
Reitse's blog 