Better performance using text-embedding-3-large?

API
1

Im working on a project using embeddings (text-ada-2)
Has anyone seen any significant jump in performance using text-embedding-3-large, specifically using more dimensions? i.e. 3,072

I wonder if using more dimensions packs the meaning of chunks better and allows for better / more effective vector search. If anyone has seen any performance boost using the new large model Id love to know.

thanks

2

We’ve seen that the new models allow concepts to be orthogonal (cosine similarity near 0) , which was a near impossibility with ada.

this allows you to outright reject documents without using an arbitrary cutoff.

Whether you need that many dimensions depends on your actual use-case.

Of course, search slows down with more dimensions, but what you can do is use the dimensions parameter to subsample your embeddings while still taking advantage of the “smarter” model.

There is a substantial difference to ADA in my opinion - but it doesn’t necessarily have anything to do with the dimensions per se.

Does that sorta answer your question?

2 Likes
3

Your entire question was exactly what I wanted to ask. Have you found out anything?

4

Thanks, yeah I think its a lot more nuanced than I originally thought.
it looks like there might be some possible gains with respect to accuracy and maybe like you said I can still see those gains by suing the new model with maybe the same # of dimensions.

I guess it depends a lot on the specific use case, will run tests next week.

Thank you for your insights

1 Like
5

@Diet’s response was quite comprehensive, Ill run tests next week and update this entry with my findings :))

6

does anyone know if there’s a performance difference between the two models while holding vector dimensionality constant?

7

I have read there is about a 4-6% performance improvement between text-embedding-3-small and large at 1536D.