It’s a mistake. I’ll fix it. I think I re-ran the query with 71 to try to pull in results that were relevant to the other poll related tables.
I deviated from the blog post’s suggestion of running all the SELECT queries with this form:
SELECT * FROM polls LIMIT 3;
I did that because I’ve got a lot of anonymized users and deleted posts in my development database. I figured I’d be supplying more coherent results this way, but I’m going to try re-doing the prompt with a fresh database and simplifying the SELECT statements.
Yeah, I’m using three users in all the examples. Their ids are 1, 2, and 121, so those values repeat a lot. I was assuming that it would be best to show coherent data. I’ll try a few different approaches and see what works best.
Another approach that’s suggested in the blog post is to limit the columns that are set in the prompt. That’s tempting, but risks introducing a lot of errors, would be difficult to maintain, etc.
The pattern I think I’m seeing is that if I start a session by asking for a complex query, ChatGPT gets confused. I get it to work through its confusion and then the results are quite good for the rest of the session. The other approach that seems to work is to start out with a simple query and work up to more complex queries from there. I’m not sure if that’s a real pattern, or if the success rate is actually more random than it seems.
As it is now, it seems like it would be useful to someone who’s already familiar with SQL and the Discourse database. I’d like to see it get to the point where it would be useful for someone who didn’t know much about either.
I’m also testing this with ChatGPT-4. It seems likely that’s going to produce better results, but maybe be less fun to use. ChatGPT-3.5 is a lot faster.