No worries, I'm pretty bad at describing the benefit of this project.
Let's say you wanted to use Gemma4B to run your workflow, however that model doesn't always do what you want it to. You could write the Margarita code to do the variable tracking, loops, etc and use the llm call where Gemma is able to work better. It's about finding the sweet spot of code vs llm agentic/tool calling.
Can you write that in js/go/etc? sure, but this lets you write it in a much simpler syntax that's much closer to markdown.
Let's say you wanted to use Gemma4B to run your workflow, however that model doesn't always do what you want it to. You could write the Margarita code to do the variable tracking, loops, etc and use the llm call where Gemma is able to work better. It's about finding the sweet spot of code vs llm agentic/tool calling.
Can you write that in js/go/etc? sure, but this lets you write it in a much simpler syntax that's much closer to markdown.
There's a few example on the website: https://www.margarita.run/custom-ralph.html https://www.margarita.run/components.html