Dimensionality is simply the number of values in a vector. An embedding model that outputs 768 numbers per input produces 768-dimensional vectors. This single number has wide-ranging effects on storage, speed, and quality throughout a vector system.
Higher dimensionality gives a model more capacity to encode subtle distinctions, which can improve accuracy — up to a point. But it also costs more: every extra dimension increases the memory needed to store each vector and the computation needed to compare vectors, and it can amplify the curse of dimensionality, where distances become less discriminating. Doubling dimensions roughly doubles storage and search cost.
Choosing dimensionality is therefore a balance. Common embedding models range from compact 384-dimension vectors to 3,072 dimensions or more. The right choice depends on your accuracy needs, latency budget, and scale; often a smaller model that runs faster and cheaper delivers nearly the same quality as a much larger one, so bigger is not automatically better.