Is there a performance difference between text-embedding-3-small @ 1536 length and text-embedding-3-large @ 1536 length?

Using kapa.ai on OpenAI Discord Prompt For the following text, summarize the idea, extract meaningful questions and give possible answers. Is there a performance difference between text-embedding-3-small @ 1536 length and text-embedding-3-large @ 1536 length? Reply Idea Summary: The user is inq…

[image] EricGT: while the text-embedding-3-small model offers increased performance over the second-generation ada embedding model. So far, that’s not my experience. I’ve found it inferior to ADA 2 in real world tests. YMMV.

I think the performance difference would be that if you requested 1536 dimentsion from text-embeddings-3-large, you would get an API error, as it is not a supported parameter. However, 1024 that is supported is essentially the first 1024 dimensions returned out of the 3k. It is hard to imagine that…

I am able to generate 1536d vectors using large

Interesting. It appears that the dimensions API parameter is just a truncater and normalizer 2048: {‘data’: [{‘embedding’: [0.018731665, -0.03743733, -0.0029751628, -0.007448469, -0.0030954042], ‘index’: 0, ‘object’: ‘embedding’}], ‘model’: ‘text-embedding-3-large’, ‘object’: ‘list’, ‘usage’: {‘p…

Yes, see: [image] It looks like 'text-embedding-3' embeddings are truncated/scaled versions from higher dim version API I was hacking around with the new embedding models and hypothesized they were all inherited from the larger dimensional version. This looks…

Yes, and to expand on my contribution to that topic… Here’s another thought: If reduced dimensions were remapped so that dimensions most relevant for obtaining semantic distinguishment were placed first to tolerate truncation, trained by extensive trials and then sorting the output order, how woul…

Given the MTEB scores in the blog post which reveal that text-embedding-3-large @ 256 length reaches a higher score than text-embedding-3-small @ 512 length, I would say it is likely that text-embedding-3-large @ 1536 length also outperforms text-embedding-3-small @ 1536 length.

Nice observation. Thank you. I’ll use large for my 1536

[image] LinqLover: Given the MTEB scores But again, that’s a case where the truncation designer can specifically target the benchmark and the dimensions that perform best on it, and where on a large parameter model, there are more to choose from. There’s no promise from OpenAI that MTEB score…

Embeddings performance difference between small vs large at 1536 dimensions?

API

LinqLover February 7, 2024, 9:20pm 7

Yes, see:

Topic		Replies	Views
Better performance using text-embedding-3-large? API embeddings	6	7297	February 11, 2025
Transitioning to the new embeddings models from ada API embeddings	8	6600	January 27, 2024
Are OpenAI text-embedding-ada-002 embedding model greater than text-embedding-3-large? Community embeddings , chatgpt , api	1	2233	February 21, 2024
Quality of embeddings using davinci-001 embeddings model vs. ada-002 model API embeddings	14	4669	December 11, 2023
Should we update to the new embeddings models? API api	4	1436	February 10, 2024

Embeddings performance difference between small vs large at 1536 dimensions?

Related topics