ai-chat.ifcopenshell.org is live!

aothmsaothms
in General

I just pushed the first version of https://ai-chat.ifcopenshell.org/ online!

This is based on bruno's remarkable ifcmcp work. But then with a web-based assistant-like interface around it. Fully browser-based. See https://github.com/IfcOpenShell/IfcOpenShell/pull/7660

Currently BYOT (bring your own token) OpenAI-only because of laziness. That's bad, but it's more of a demonstrator.

The usefulness you can get out of this even today is really amazing. If you haven't already tried the full ifcmcp I really recommend to have a look.

carlopavbrunopostlezoomertheoryshawMassimosemhustejmyoder89walpafalken10vdlMoultand 4 others.

Comments

  • theoryshawtheoryshaw

    very cool!
    Possible to add a Claude key too?

  • aothmsaothms

    Maybe ask Claude to add it: https://github.com/IfcOpenShell/IfcOpenShell/tree/v0.8.0/src/ifcchat cannot test myself. It's not only the token, it's also the API calls: https://github.com/IfcOpenShell/IfcOpenShell/blob/fb81c88a5f03f2dc4a11d5c93aadb990eac06981/src/ifcchat/app.js#L162

  • theoryshawtheoryshaw

    Cool, will try.

  • myoder89myoder89

    Claude Code got it set up for me pretty quickly. Playing around with it now!

  • theoryshawtheoryshaw

    @myoder89 cool, can you push the change as a PR? Thanks.

  • theoryshawtheoryshaw
    edited April 2

    someone provided OpenRouter support.
    That's cool.
    https://github.com/IfcOpenShell/IfcOpenShell/pull/7896

  • myoder89myoder89

    Sorry for the wait everyone. PR for addition of Claude is here, let me know if you have any issues:
    https://github.com/IfcOpenShell/IfcOpenShell/pull/7897

    brunopostletheoryshaw
  • aothmsaothms

    Thanks all. Really looks a bit better with provider selection.

  • theoryshawtheoryshaw
    edited April 3

    I have credits, but Claude cried when I fed the following prompt, on the following model.
    https://hub.openingdesign.com/OpeningDesign/Bonsai_Tutorials/src/branch/main/_Model/House.ifc
    It even cried when I shortened the prompt to 'just make the basement walls taller'. 

    Can you make the basement walls 1 ft taller?  That is, move them down 1ft and extend them up 1 foot.  Can you also move the footers, at the bottom of that wall, down 1ft.  Can you also move all the bleeder pipes, on either side of the footer, down 1ft as well?  Can you also move the basement slab down 1ft, and any associated thickened slabs. 

Can you adjust the basement stair, and the associated stringers, to this new basement height now.  The riser should not exceed 8".

On all the ifcannotation drawings, can you move all the associated annotation to these objects down as well? And can you extend any ifcannotation dimensions, as well?
  • aothmsaothms
    edited April 3

    30_000 tokens per minute sounds like something you would burn through pretty quickly, but I also don't see a real monster in terms of output. If it's really minute based then I think the most likely way out is to introduce a small delay in terms of the tool call iterations so that you spread out the execution over a longer period of time.

    I don't have much experience building these kind of apps. Curious what others say. The fact that you send the full history for every message does sound a bit awful in terms of non-linearity.

    Edit: what I did is capture the body before being sent. That was about 1000 words, so approx 3000 tokens (?). But that's a single message. With all these tool calls + full history that quickly adds up.

    theoryshaw
  • myoder89myoder89

    I had Claude review and it pointed out a few things that could improve the token usage per prompt. I tested locally and it seems to have greatly improved what ChatIOS here can handle. Here's the link if folks want to test it out. Fair warning, this was a 100% "vibe" on my part.
    https://github.com/IfcOpenShell/IfcOpenShell/pull/7904

  • aothmsaothms

    Thanks!

    Fair warning, this was a 100% "vibe" on my part.

    Well, at worst, that moves the total vibe percentage up from 95 to 96% ;)

    theoryshaw
  • theoryshawtheoryshaw

    Assuming these latest changes were applied, i still get the following, with the same prompt and file, as mentioned above.

  • aothmsaothms

    There is now also compaction, i.e using the llm to summarise the message trail to reduce tokens and some client-side throttling / rate limiting to keep track of the estimated amount of input tokens per minute bucket and if at limit defer to the next minute.

    I'm curious if you get better results now @theoryshaw your case still appears to be somewhat advanced, but it would be interesting to see what's needed to get this to work:

    • fine tune the system prompt with instructions because it seems to get stuck in similar occasions?
    • but also tweaking the api or its outputs to be at the right abstraction level.

    It's probably better if you use ifcmcp directly though because I imagine the kind of compaction and throttling is built in if you use their official client.

  • theoryshawtheoryshaw

    Yeah, unfortunately, got the same error, even with...
    Can you make the basement walls 1 ft taller? That is, move them down 1ft and extend them up 1 foot.

Sign In or Register to comment.