What is Embedding Dimensionality?
Also called vector dimensions.
Embedding dimensionality is the number of values in each vector produced by an embedding model, commonly a few hundred to a few thousand. It is fixed by the model, and every vector in one collection must share it. Dimensionality drives storage size, comparison cost, and, up to a point, how much meaning a vector can carry.
Storage and compute scale directly with dimension count. A collection of ten million vectors at 1,536 dimensions in 32-bit floats occupies roughly sixty gigabytes before any index overhead; halving the dimension roughly halves that. Distance computations scale the same way, so dimensionality shows up in both the monthly storage bill and the per-query latency of any similarity search.
More dimensions do not automatically mean better retrieval. Beyond the point where a model has captured the distinctions that matter for a task, extra dimensions add cost and can dilute distance contrasts, an effect loosely related to the concentration of distances in high dimensional spaces. Whether a larger vector helps is an empirical question answered by evaluating retrieval quality, not by comparing dimension counts.
Some newer embedding models are trained so that a vector can be truncated to a shorter prefix and still work, a property usually described as Matryoshka representation learning. This lets one model serve a cheap low dimension index for a first pass and a full length vector for reranking. Support is not universal, and truncating a model not trained for it degrades quality unpredictably.
The practical constraint is that dimensionality is part of a collection's contract. Vectors from models with different dimensions cannot be compared, so changing the embedding model usually means creating a new index and re-embedding the whole corpus. Recording the model identity and dimension alongside every stored vector is what makes that migration traceable later.
Key points
- Set by the embedding model, identical across a collection
- Storage and distance cost scale with dimension count
- Higher dimensions do not guarantee better retrieval quality
- Some models allow safe truncation to shorter prefixes
- Changing dimensions requires re-embedding the whole corpus
In practice
A team compares a 768-dimension and a 3,072-dimension model on the same two thousand labeled question and passage pairs. The larger model improves top-five hit rate from 0.81 to 0.84 but quadruples index memory and doubles query latency. They keep the smaller model for first-pass retrieval and spend the saved budget on a reranking step, which lifts hit rate further than the dimension increase did.