Voice of the DBA

Tag: AI

Prompt AI helping with Auditing

I had a conversation with a customer asking this question: how can I tell who called a stored procedure so I can audit the action?

I decided to see if Prompt could help me here.

This is part of a series of experiments with AI systems. This is also part of a series of posts on SQL Prompt. You can see all my posts on SQL Prompt under that tag.

The Setup

I had this code, which is a simple stored procedure. I sketched this out, but I’m looking for a function to add to my code that helps me audit the caller.

I opened the Prompt AI (ALT+Z) and asked the question you see in the image. I also asked to assign it to the variable, but didn’t exactly specify everything.

In the image below, you can see the code that changed. Prompt AI added the System_User function, which is exactly what I wanted.

I accepted this and then asked to change this to the user instead of the login.

Again, the function I wanted was added: CURRENT_USER.

I know what I needed, or I had an idea. However, if I wasn’t sure what function to use or if there was one, this is handy. I’m working inline, rather than going to a browser and hitting Books Online or Googling for the answer. I can keep coding or looking at my code as the AI runs.

If I ask for an explanation, I get one, which helps me judge if this is actually the code I want.

If you have SQL Prompt, get Prompt AI a try by opening it with ALT+Z.

If you haven’t tried SQL Prompt, download the eval and give it a try. I think you’ll find this is one of the best tools to increase your productivity writing SQL.

Video Walkthrough

I show this live in the video below.

October 15, 2025
Adding a Local Model to Ollama through the GUI

I saw some good reviews of the small gemma3 model in a few places and wanted to try it locally. This

If you want to get started, read my post on setting up a Local LLM. This post gives an alternative to connecting to the container from the CLI and running a command.

This is part of a series of experiments with AI systems.

Adding a Model

Using the Ollama-WebUI server, it’s easy to add models. I wrote about setting up the UI and I’m running that as my interface to the local model. I created a new container for this post, so when I see the interface, there are no models to choose from. You can see this below.

I need a model because if I try to ask a question, I get an error that the model is not selected. You can see the error at the top in this image and the query at the bottom.

If I click the gear icon by the model drop down, I get a list of settings. One of these is “Models”, which is third on the left.

I click Models and get this list. As you can see I have no models. Notice I need a model name to enter. However, where the cursor is below, there is a “click here” for a list of models.

This opens the Ollama site at https://ollama.com/library

If I scroll down, I see the gemma3 model. If I click this I see a few different ones. I am not a big fan of latest, but this is a test. I decided to just grab it since I’m not programming things.

If I enter this in my WebUI and click download, it starts downloading.

Progress is shown, and it’s not quick. I went on to other work and this is the view about 15 minutes later.

Once this is complete, I close settings and see my model in the list. Since this is the only model for this container, I’ll click “set as default”.

Now I can get my fun fact:

Now I have a safe, secure, local model to use. If you want to see this run in real-time, check out this video:

October 13, 2025
AI Model Size, Parameters and Download
At SQL Saturday Boston 2025, I gave a presentation on local LLMs and there was a great question that I wasn’t sure about. Someone asked about the download size vs the model size as listed on Hugging Face (or elsewhere). It was a good questions, and I assumed the parameter size (258M, 7B, etc.) relates to the download size or size on disk.

I did a few searches and this was a great article on LinkedIn, called The Intergalactic Guide to LLM Parameter Sizes. In it, there’s a quick guide:
- Tiny (1-3B parameters): 1-2GB on disk
- Small (4-8B parameters): 3-5GB
- Medium (10-15B): 8-15GB
- Large (30-70B): 20-40GB
- Enormous (100B-200B): 60-150GB
- Apocalypse-Inducing (500B+): 300GB+
That’s a good rule of thumb, and a good thing to know. Now, the storage isn’t that important, but the power consumption is something.

I have a laptop which has these specs:
- 12 core Ultra 7 Intel CPU, 2.1GHz (base)
- 32GM RAM
I asked a simple question of my local model (Ollama, mistral 7b), so a small model, and I see this in task manager

The power consumption in the guide says a small desk fan and any modern laptop can run it. I was playing with this on a plane and could see my battery going down. I don’t think that the storage is a problem on most modern laptops/desktops, though certainly you want a T/S/M/L model running. The power, however, can be a challenge, especially if you don’t have a good GPU. My laptop has an Intel NPU (AI Boost) designed to help AI stuff though I don’t know if the container uses it. I also have a GPU from Intel, but still, this GenAI stuff uses power.

Be careful on a laptop that isn’t plugged in if you need to get a lot of work done.
October 1, 2025
Choosing an AI in Sept 2025

Some of you out there are positive about AI and looking to use them. Some of you aren’t too thrilled with the tools and might avoid using them. I think that not learning to harness the power of an AI is a mistake. This technology is going to change the world in many ways and you need to learn how it can help you.

You also should learn where the sharp edges are as there are some very, very sharp edges.

Which AI should you use? That’s an interesting question. If you listened to my podcast interview with Tom Hodgson (YouTube, Spotify, Apple), he prefers Gemini as a developer at Redgate. I think that’s one of the more popular models to choose from as a developer. I see lots of people writing about their experiences using Gemini to code.

It’s recommended (with others) in this article on choosing models, which is a topic that I think is important to keep an eye on. We often like to build some skill and trust with a tool, developing muscle memory that helps us work efficiently with the tool. In the AI world, things are changing constantly, though not so much that I would recommend jumping on the latest release of a new model.

Instead, I’d recommend you stick with a tool and, perhaps every 3-4 months, you experiment with another model and see if you like it. Perhaps send some duplicate queries to your current tool and a second one. As much as we’d like to say this one LLM is best, there is some personal preference we will have for different tools. I find LLMs to be non-deterministic, so you want to work with a tool in which you have some trust.

The current top models that most people look at are the OpenAI gpt models, the Claude models, and the Gemini ones. It is hard to keep track of these, so I’ll try to simply them.

For many of us in the Microsoft space, we are used to OpenAI and the ChatGPT models. There is the gpt-4o and gpt-5o models in use by many people. I haven’t found these as useful to me for SQL coding. The way they respond slightly grates on me, so I have learned to check in various tools to switch the mode away from gpt-40 (often the default). They might work well for other code, but they seem to lean towards MySQL, which annoys me.

Claude (from Anthropic) is my preferred model. I usually choose a Sonnet model (3.7 or 4) and have them work well. I like the structure of Sonnet output and find it helps me get database work done pretty well. It’s not perfect, and it’s certainly not always correct, but it’s good. I’ve had it solve a number of problems, and it’s the one I often use in my AI Experiments posts.

Gemini (from Google) is the one I’ve used the least, but it’s the one several developer friends like. It’s on my list to experiment with this one a bit and see how it works. There are various types of v2.5 models here to choose from.

Depending on what types of work you do, you might like other models. I primarily deal with text, so I haven’t spent much time with models that work with audio/video/pictures, though my wife has been playing with AI video editing. She doesn’t know which model she used, but she appreciated the LLM help in producing a sale video for a horse recently. It was way better than anything she had done on her own.

I’ve tried Perplexity (sonar), and very lightly, grok, but they haven’t attracted me for some reason. I tried a local DeepSeek R1 model, but it wasn’t smart enough (or quick enough) to get much done. If you’re in doubt, just ask the AI which model is running. Most of them let you switch models easily, so give it a try.

And let me know what works for you and why. I have found different models working differently, so pay attention and share what you learn.

Steve Jones

Listen to the podcast at Libsyn, Spotify, or iTunes.

Note, podcasts are only available for a limited time online.

September 29, 2025