The embedding dimension is the number of values in each vector that an embedding model produces. It is fixed by the model: a model described as 768-dimensional always outputs vectors of exactly 768 numbers, regardless of whether the input is a word or a paragraph.
Dimension is a central trade-off in choosing a model. More dimensions give the model more room to capture subtle distinctions, which can raise accuracy, but they also increase the memory each vector consumes and the time each comparison takes. A larger dimension multiplies storage and search costs across the entire database.
Common models range from compact 384-dimension embeddings to 1,536, 3,072, or higher. Higher is not automatically better — beyond a point, extra dimensions may add noise rather than useful signal, and a smaller, faster model often delivers nearly the same quality at a fraction of the cost. Some newer models even support adjustable dimensions, letting you truncate vectors to trade a little accuracy for major savings in storage and speed.