Nick Crews

Yes I think this would be good. open questions? - how to enable/disable. I would say if debug mode is on then just include it. Could also adjust `.debug` to...

> If you only have the product name, i suppose you could consider trying to get the embeddings down to a small array and splitting out to columns (embedding_1, embedding_2...

What is preventing us from implementing the cosine similarity in SQL, so it works on all backends? Unless I'm missing something it wouldn't be that hard to generate some SQL...

That's true. For that case, each backend might have slightly different syntax, so we should make it possible for each backend to override with their own implementation. But it can...

I think we could keep the same API as the function I wrote above, but if you supply a single string (instead of Iterable[str] like I have there), then it...

Native implementations (eg the one in scala above) are going to scale O(length of vector) just like the above SQL. You are always gonna need the same number of basic...

Y'all should look at how Ibis does this with generic Operations (eg Join, OrderBy, ArrayAgg, etc), with specific backends then registering backend-specific implementations. They have some very mature and clean...

It looks like your screenshots are from the EM training session charts. I think those charts are adequate as they are, I just need to see how M and U...

That sounds good. I would lean towards the default being `include_mu=True`, since that is what I personally will always be doing, and I don't feel the charts are particularly visually...

I'm sorry, I don't think I understand your problem still. What would your code look like if this feature was implemented? Is this a jinja2 problem, or a staticjinja problem?...