Cool! I'm working on something similar (https://app.embedsheet.ai, a bit buggy and in-dev, the paid stuff is just placeholders) but yours feels much more responsive and smooth than mine, it's very impressive!
I notice that one of the big things with using LLMs for something like this is making that each cell input/output has a corresponding cell reference (e.g. A1:"data in a cell"). I was originally doing it with CSV data inside XML tags for each block of data
e.g:
<data range="A1:B2">
1,2
3,4
</data>
Which works for very small blocks but since LLMs are so terrible at counting it would begin to make mistakes after anything more than a very small range. I think the killer feature for something like this would be letting the LLM try to come up with JS-based custom functions so that it's not limited to builtin formulas.
Cool! I'm working on something similar (https://app.embedsheet.ai, a bit buggy and in-dev, the paid stuff is just placeholders) but yours feels much more responsive and smooth than mine, it's very impressive!
I notice that one of the big things with using LLMs for something like this is making that each cell input/output has a corresponding cell reference (e.g. A1:"data in a cell"). I was originally doing it with CSV data inside XML tags for each block of data e.g:
<data range="A1:B2"> 1,2 3,4 </data>
Which works for very small blocks but since LLMs are so terrible at counting it would begin to make mistakes after anything more than a very small range. I think the killer feature for something like this would be letting the LLM try to come up with JS-based custom functions so that it's not limited to builtin formulas.
Very nice! thats a great idea. I was thinking xml tag by tag is good because it creates a lovely streaming experience of cells filled in one by one
With IronCalc!!! Soo incredibly happy to see this
(BTW I just saw this two minutes after submitting by complete chance)
Oh amazing! Ironcalc was awesome to plug in.
Was reading issues you reviewed while building it.