Hi, Yesterdays awesome demo of Scrappy during the FoC meetup and the small "one more thing"-moment about integrating Scrappy with AI/LLMs, reminded me of posting something here about 2 of my own recent new side projects: - https://ocif-generator.vercel.app : generate OCIF files and view them as JSON/SVG/React-flow - https://app.prompttoform.ai : generate complex forms and play with them in the preview and also view its structure in the flow tab using react-flow. The latter is especially handy when the form contains multiple steps and has decisions in them (because you can build decision tree's with this generator). Both tools use structured output using json schema's for the llm's, which allows for a lot of control instead of just prompting and trying to get structured output via a text prompt. Both of these tools were coded using a lot of AI as well (using Cursor). The OCIF-generator was demoed yesterday at localfirst-conf by one of the OCWG core members @Jess Martin My plan is to integrate both of the above with my own infinte canvas visual programming system: https://codeflowcanvas.io. Check https://canvasprotocol.org for more info about OCIF.

Cool! I haven't been following LLM APIs recently so maybe you'll know... Is there a way to make them ingest/produce embeddings for graph vertices or edges? IIRC just before LLM explosion transformers were making some noise because of their ability to process graph-shaped data natively.

You can create graphs with llm's, thats what I do using that small generator.. but the layout is done without llms in that generator .. I am experimenting with different libraries like cytoscape and d3 forced layout.

It's called graphical neutral network. They attach feature vectors to vertices and/or edges and can be used to build all sorts of graph processing primitives. They can use transformers under the hood so I guess that somebody already had to merge them with LLMs. Mixed modality text/graph data is possible. I was just wondering if there are any APIs that allow graphs as input or output (similar to how some LLMs allow audio or images).

I dont know.. and do you mean graph neural networks?

Ah yes

Thanks for bringing this up! I am going to dive more into GNNs , it sounds like a powerful combination with LLMs