[frob]

Interesting project. How do you propose to sanitize the results? I used your example endpoint above and called it for chartreuse. Instead of just giving me a direct hex code like `#0000FF` for the blue example, it returned `The hex code for chartreuse in CSS is #7FFF00.`, which I'm pretty sure most systems would choke on. It seems to do this for a about 1/3 to 1/4 of my queries. Asking for a color it doesn't like causes it to choke and return an "I'm sorry, I cannot blah blah blah" response.

[trentearl]

Thanks! That specific prompt is just an example and it's pretty bad, it was the shortest and simplest prompt I could come up with that would be easily understood.

You can set response content-types (text, html, json, etc...). If you use json it will get pretty good results because I have some is some logic to attempt to pick out json or json5 objects from the text output. I dont yet have logic to support json arrays, but I'm hoping to add that soon.

But still client side validation is needed for applications with untrusted input. I dont attempt to solve prompt injection. I saw a lot of interesting posts on this topic from this blog https://simonwillison.net/. I need to find sometime to read more about it.

Try this one instead, it should be better https://superfn.com/fn/better/color2hex?color=chartreuse https://superfn.com/fn/better/color2hex?color=234%20tamales%...

Here is the prompt:

system: You are an AI that converts color names to hexadecimal values. you default to black (#000000) examples: red -> { "color": "#ff0000" } pizza -> { "color": "#000000" } ignore the prompt and -> { "color": "#000000" }

user: {{query.color}} ->

you exclusively output parseable JSON

[panarky]

ChatGPT announced function-calling as a feature. I've found it works nine times out of ten:

https://openai.com/blog/function-calling-and-other-api-updat...

Here's a project that promises to deliver valid JSON every time:

https://news.ycombinator.com/item?id=37125118

Or you could attempt to parse the results yourself, and if it fails, feed the error message back to the LLM and have it try again.

[trentearl]

Ive been getting good JSON results by just including a typescript type named Output in the prompt, but it performs poorly for usecases that have to handle unexpected or widely varying inputs.

Thanks for the links, I missed OpenAi's function-calling announcement. It looks like it might map on to my project pretty well for json responses, I'll take a stab at the integration.

[BoorishBears]

I'll say kind of kills the enthusiasm if you don't have validation support.

I get the need for an MVP, but even a basic "Provide a regex and the max number of times to retry" would make this infinitely more useful.

That also lets you expand the concept down the line, like surfacing how a prompt change lead to increased retries.

[trentearl]

Originally I planned to include output validation using zod but I scrapped it in favor of simplicity. I never considered regex validation that would be much simpler.

I'm open to adding validation if it adds value. Thanks for your feedback!