Dennizz

If you mean the common OpenAI API key, I'm not sure how that's different in regards to feeling secure compared to separate keys. Would you mind elaborating further, please?

To clarify, the common API key is known only by my server, not by the app!

Not yet but I have it in my list of TODOs!

I just pushed a new release. It should work now!

To install the new version, just restart the app. It should then prompt you to install the update.

Typescript and Rust, with SvelteKit and Tauri.

That's fair. The "no-tricks pricing" wording was taken 1-to-1 from the UI template I used :)

I'll update the copy, and try to explain the pricing model more clearly.

Also a good idea to make it possible to use one's own API key. I'll consider that.

It's GPT-3.5 only so far but I might add additional models at some point.

Right now it's pretty simple: First I drop the table/column information if the token limit gets exceeded otherwise (this is shown as a warning in the UI). Then if the limit still gets exceeded, the I show an error to the user.

I plan to make this process more smart eventually.

I'd like to use it eventually but as of now API usage is too limited.

I've had mostly the same experience as calmoo.

And when it does go wrong, I've had good success telling it what it did do wrong. It will usually correct itself then.

Correct, you probably should know SQL.

In my experience though it is often a significant time saver if I only have to review the output of the AI vs. coming up with everything myself.

Re: ambiguities: You are right, the AI won't always be able to infer the correct fields to use. In such cases though it is often enough to help the AI out a little. E.g. by saying "information X is stored in field Y of table Z".

Even if you have to do that, you can still save time and more importantly mental effort by letting AI help you, compared to writing all the SQL yourself.

Re: iterative process: You can simply send a follow up message saying "You did X wrong". It is usually happy to correct itself.

No, currently it doesn't.

This is a case where it could help to give the AI some example values of the actual data. The same applies for e.g. JSON fields.

I've been thinking about ways to make it possible for users to easily provide examples to the AI. Essentially that would mean selectively giving access for specific rows/columns to the AI.

Interesting idea. Do you have some examples of how you would like to use that, or maybe even an app where you'd like to integrate it?