Random 384 dimension embeddings from text-embedding-3-small

API
1

I am getting errors again today… same error: text-embedding-3-small and an embedding vector of size 384 (instead of 1536 as expected)

5 Likes
Embedding Model Outage: text-embedding-3-small-api-ev3 model name with all 0 values
2

I am facing similar issue today while using text-embedding-3-small the model is randomly generating 384 dimension embeddings instead of 1536 dimension.

Also an observation, its not consistently generating 384 dimension embeddings sometime it does generate 1536 and sometimes it generate 384 dimesion embeddings.

2 Likes
3

Hi and welcome to the community,

I was not able to reproduce this. In my tests, text-embedding-3-small behaved as expected and consistently returned vectors with the correct number of dimensions.

I have created a new topic for these two reports and will monitor the situation.

4

just to add to this I saw a similar thing happen at 5:14:30am GMT across multiple services calling the same function for embeddings

5

Same issue! Even after explicitly setting the dimensions, it keeps happening. Has anyone found a solution?

6

Solution, if you’ve got 2GB of GPU RAM and want your embeddings to work for you forever at half the storage:

Rank 8 and 10, under 2GB: jina-embeddings and stella) across columns like zero-shot, memory use, parameters, embedding dimensions, max tokens, and mean scores.
Rank 8 and 10, under 2GB: jina-embeddings and stella) across columns like zero-shot, memory use, parameters, embedding dimensions, max tokens, and mean scores.1237×257 56.4 KB

"embeddinggemma-300m" ranked 14 with 99% zero-shot, memory usage 1155, and other metrics including max tokens 2048 and mean scores around 61.15 and 54.31. (Captioned by AI)

Not:

A leaderboard-style table row shows rank 44 for the model “text-embedding-3-small,” with a zero-shot value marked NA and columns listing embedding dimension 1536, max tokens 8191, and a mean task score of 46.85. (Captioned by AI)

The last one: run thousands at once without thinking about the hosting or utilization.
When it works.

7

Hi and welcome to the Community!

I reran my tests but couldn’t reproduce the same behavior. We can still escalate this to the team for review, but since it isn’t consistently reproducible, please capture the x-request-id from any request where the issue occurs.

1 Like