Skip to content
BENIBIX Learn

Embedding Dimensions

Also known as Vector Dimensions, Embedding Size, Vector Length

How many numbers each embedding holds. More can mean finer meaning, but more cost.

Think of it like

It is like describing a fruit with a list of scores. More scores tell more, but take more paper.

Example

One popular embedding model gives 1,536 numbers per text by default, and its larger sibling gives 3,072. Emma asks for 512 to save memory, and search stays almost as good.

Why it matters

Dimensions set the size of your vector database and its bill. Halving them can halve the storage.

Definition

Embedding dimensions are how many numbers are in each embedding. Each number scores the text on one hidden feature the model learned. A small model might use 384 numbers, and a large one a few thousand. More numbers can hold finer meaning. They also take more memory and make each search slower. Every vector in one index must have the same number of dimensions.
The count of boxes in a row is the number of dimensions.

Common mistake

Mixing embeddings from two models in one index. Their numbers mean different things, even with the same count.

Key takeaways

  1. 1
    One number per feature. Each scores a hidden trait.
  2. 2
    More is not free. Bigger vectors cost memory and speed.
  3. 3
    Match them. One index, one size, one model.