Which embedding dimension should I use for 2M short product titles?

We are embedding about two million product titles (5–15 words each) for semantic search. vx-embed-3 supports 256, 1024 and 3072 dimensions. Storage is a concern — what do people actually see in recall between them?

We tested on 500k SKU titles. Recall@10: 256 → 0.81, 1024 → 0.86, 3072 → 0.87. The jump from 1024 to 3072 was not worth 3× the storage for us.

Short titles do not carry enough information to fill 3072 dimensions. I would start at 1024 and only go up if you can measure a gap on your queries.

Ran a quick test: 1024 it is. 0.85 recall on our set, 8 GB of vectors instead of 24. Thanks!