So I sense the Wappler AI functionality currently lets the AI write your code.
I have an application where I'd like AI to actually be the code.
I want to create an appointment booking system... and I'd like it to link to an AI agent which is passed calendar information about when the practitioner is free, a load of information about appoinment types, and then have it output an HTML based list of possible appointment dates which is inserted on the fly into my existing booking form structure.
I don't know how to help in regards to AI, but I just wanted you to know AI isn't strictly required for your particular challenge, this sounds like a computer science business problem. But you have a good question about feeding inputs to AI and then getting back data in a specifically-formatter manner
For AI you could explore stuff like n8n, and use API Actions to talk to n8n. Maybe Hyperbytes might have some ideas if you prefer native Wappler, he did some custom extensions for NodeJS AI-related.
Outside AI, you'd store the different appointment types and their suggested timeframes, the allowed booking intervals... I'm not totally sure on the DB design, we'd need to research... I can see how using an LLM for event handling could skip some of the DB formalities by adopting natural language instead of formally defined structures and records
I'm doing something kind of similar. I input live data taken from SC actions on the backend and prompt the AI to write specific job ads. Its highly effectvie and does a fantastic job.
Im using the excellent extension from @Hyperbytes :
The training info prompt looks like this (this is getting honed all the time and is getting more extensive after each iteration):
You are an AI assistant providing help to write persuasive job advertisements for for '+query.Company+'. start date is '+$_GET.startdate+' job title is '+$_GET.prompttitle+' job location is '+$_GET.country+' include "(View on Map)" use this map URL in the View on Map text: '+MapLink+' make it open in a new window. The company website URL is: '+query.CompWebsite+' A brief description of the company is: '+query.CompComments+' the companies country is: '+query.CompCountry+' format your replies correctly using paragraphs, bullet points and headings where necessary. important: Include hyperlinks for all links given. Use this link to '+query.Company+' profile: '+CompanyDetailsLink+' whenever you can. Use bullet point lists for the job title, start date and location Do not use any headings larger than
Do not underline headings. Do not include backticks or . At the end write in bold and on a new line: To apply for this position, please use the "Apply Now" button on this page. Aim for a word count of approximately 150-320 words. Use HTML instead of Markdown in the response.
Yes, we do this for certain non-mission-critical areas. You'll want to multi-shot train the model using whatever dashboard your model of choice offers, so that you give it tons of examples of the exact code structure outputs you expect, along with fail-cases. (Or slam it all into a large prompt, but token costs are sometimes cheaper with pretrained/fine-tuned models, depending on your expected token levels).
Bear in mind, you'll be severely 3rd party dependent, and face uptime/SLA risk.
Me: patiently waiting in the corner for server GPU's to get cheaper to be able for us all to standup open source blackboxes...
