Skip to main content

Using Gen AI to write a fairly simple SQL query

So I wanted to see if I could test the different Gen AI models that are out there and get them to write a relatively simple SQL query. Basically select against my table, as detailed in the prompts to Gen AI, and produce a list of the fastest 1000 times at an event (that takes place weekly) and provide the times and names of the athletes that ran said times. Note that although I say view a lot I mean query because what are views if not stored queries anyway and I am using this in my DB as a view. 

Winner: Copilot

The original view can be seen below: 


So it is a fairly simple view with some logic in it to through some spanners in the works. The question is with the table definition and some explanation can the Gen AI platforms recreate a working version of the above view? 

The initial Prompt: 

I can't find a good way to format and embed my whole chats with the AI tools so I will work with what I have. Here is my original prompt that I used to get a starting point. 



Gemini: 

Overall this chat isn't very long and the first attempt was pretty close, it just missed the floor function around the minutes column and some quotes so I asked it to add those. 

"Please add a floor function to the fastest_time / 60 AS mins column Also the cte needs the double quotes around theparkrunID and Name columns. "

And here is the Gemini view, which I tested and it produced the right results.

The Gemini View:

Here

Copilot:

Next up I tried CoPilot. Copilot gave me a working piece of SQL out of the date, I did have to ask to expand the comments at the top and to add minute / seconds to the output but I like the way it did it. Overall was very impressed with it. No problems at all.


Copilot View:



ChatGPT:

Literally identical story to CoPilot with a slightly different output. I got a working view first time but didn't have the additional columns that were useful and it didn't get the comments right at the top and when I asked it to add them I think it went overboard! Sorry for the formatting it won't let me fix it!

ChatGPT View:



Summary:

Gemini - Original Prompt and 1 correction and gave a working view. Provided me with some nice to haves that the others didn't.
Copilot - Original Prompt worked and minor additions asked for and gave a working view.
ChatGPT - as with CoPilot

Think that overall the winner is Copilot, the final solution it gave me worked and it put in nice comments. I personally would then go with Gemini, even though it require a correction to add the floor that the others didn't it gave the view most closely matching my original and got the comments right. This leaves ChatGPT in last. Honestly however the experience with all 3 was very similar and they all produced more or less what I wanted within a handful of prompts. I wonder how they will cope with more complex data models.



Labels:

Comments

Post a Comment

Popular posts from this blog

Gen AI news 29-04-2024

Here are some recent updates and insights related to Generative AI (gen AI) : Enterprise Hits and Misses - Robotics and Gen AI Converge : This article discusses the convergence of robotics and generative AI. It explores breakthroughs needed in the field, the FTC’s policy change regarding non-competes, and the impact on AI model sizes for enterprises 1 . Read more All You Need To Know About The Upcoming AI-Powered OLED iPad Pro : This piece provides a summary of rumors surrounding the next-gen AI-fused OLED iPad Pro, powered by the new Apple M4 chip 2 . Read more Delivering on the Promise of Gen AI : New Electronics reflects on NVIDIA GTC and key announcements that contribute to delivering on the promises made for generative AI 3 . Read more The Future of Generative AI - An Early View in 15 Charts (McKinsey): Since the release of ChatGPT in November 2022, generative AI has been making headlines. McKinsey research estimates that gen AI features could add up to $4.4 trillion to the globa...
Post a Comment
Read more

Keboola Flows

Really finding Keboola was the thing that kickstarted this project otherwise I would be trying to build custom code on a python cloud server and building everything from scratch.  In Keboola you build you data sources and destinations using connection details which is fairly simple and something I will likely cover in another post, same goes for transformations etc. Here though I am going to discuss Flows, this is where you bring everything together. On my free account there are some limitations.  My easiest flow is very basic:  Pull parkrun results e-mail from Gmail to Google Sheets (actually done by Zap not Keboola).  Keboola will, as often as I like, in this case once a week, pull the data from the sheet into its storage.  It will then transfer this to the target database. Currently I have this setup to be MySQL database but I can and might expand that to the Snowflake instance within Keboola.  I then, outside of Keboola, connect to the MySQL database f...
Post a Comment
Read more

Snowflake Scripting - SQL Cursors

Snowflake scripting in SQL seems to be in preview and I have decided to have a play with it. Given how new it is there is limited documentation so I am using a combination of what I can find on the Snowflake site and the odd blog that has been written about it. There appear to be a few quirks, at least when compared to Oracle PL/SQL (though that has been round for years). How many of these are intentional and how many are things to be ironed out I don't know. You can see the procedure I have created it:  Accepts an id as a parameter  Creates a result set selecting from a table, using the parameter as a filter Loads the results set into a cursor.  Loops through the cursor loading the id in the cursor into variable Calls procedure passing in the variable as the parameter.  Then as a proof of concept I tried the Snowflake feature of allowing declaration of variables within the main start and end block. 
Post a Comment
Read more