web2050 Used AI

updated to hc/ai SSE spec, works without #24 being merged. SSL ERROR FIXED !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

SSL is confirmed not fixed, I did change it to use the new API proposal to avoid some SSL issues by excluding the <think> data from the response. It can generate full small pages now, I also prompted it to output minified output .

finally got it running on nest again! SSL is an issue which will not be fixed by HC, so i can't fix it :(

worked around SSL error by using /no_think in prompts. This does not work every time.

I ACTUALLY GOT IT TO WORK!!! STREAMING XML IS REAL!

I am working on implementing a new XML parser to fix the bugs with the old one, I think it works?

Worked on fixing the XML parser. Figured out hackclub's SSL is all borked and that was why the AI was giving half-complete responses. Cannot be fixed by me, submitted a report!

Worked on fixing the XML streaming for the new ai.hackclub.com model.

improved index load times by 5x by omitting tw classes from each <li> and instead emitting them before with a <style> tag! This also fixes some bugs w/ the generator. I fixed a bug where pages would not generate the <code> tags properly to fmt it to output along w/ thonking data. Screenshot is 3.7MB and doesn't even include the whole list :O

Redesigned the logo because the AI one sucked.

I removed rust-ls then added it back because it broke the completions endpoint :skull: HTTP/2 is now removed as caddy handles that. AI Also made a final logo!

fixed some generation things in the prompt, the model is more effective at linking pages together now. Might add external images soon, or even external JS. Still deciding how bad this would be.

Manually cleaned up some NSFW before the filters, people are too fucking odd for my service.

I made the 64 URL length limit now 72. I also migrated to HTTP/2 and made agar.io!

I rewrote the backend with axum, a more maintained and newer framework than rocket. I hope this speeds things up a bit!

Implemented a new ripgrep powered search function to search the index by path name as well as file content. Fixed the CONTENTREJECTED bug, it now properly shows CONTENTREJECTED when the content was rejected instead of outputting nothing.

Implemented a new streaming XML parser to take output in the form of <code>...</code>. It does not return any data until the first code tag and only stops when the last one is present, not yielding any of the tags.

I greatly improved the system prompt and the AI's ability to make pages link to other pages, now every button link and menu has a point! No more href='#'s

I also added a hard path limit of 64 characters in order to prevent platform abuse.

Added a new landing page, as well as more AI improvements.

improved system prompt, uses 3rd person augmentation, offloads generation to another thread so closing the tab does not stop it, give the AI the current date. Added experimental sandboxed support for JS, let's hope my users don't do anything horrible with this one.... again....

Added support for tailwindcss! Implemented content filters through a CSP and the nest code of conduct. The AI has been prompted to not produce any harmful content, in addition it can no longer link to external assets. I may change this to allow images.

I started work on the initial version, currently it can generate entire webpages by domain and it can write to the filesystem. It can read files that relate to the current work, ie it can make an index.html, then write CSS for it based off of the index.html. The browser automagically requests linked CSS which gets put into a generation queue.