A vector that stands in for meaning
An embedding is a fixed-length list of numbers, a vector, produced by a neural network so that the geometric position of that vector captures something about the meaning of the thing it represents. Feed the word 'ocean' into an embedding model and you get back something like 768 or 1536 floating point numbers. On their own, none of those numbers means anything to a human. What matters is the relationship between vectors: an embedding model is trained so that inputs with similar meaning produce vectors that land close together in that high-dimensional space, and inputs with unrelated meaning land far apart. 'Ocean' and 'sea' should be neighbors. 'Ocean' and 'spreadsheet' should not be.
This works for more than single words. Sentences, paragraphs, images, audio clips, and even user behavior sequences can all be embedded, as long as there's a model trained to map that kind of input into a shared vector space. The model learns this mapping from data, typically by being trained on tasks where it has to pull related pairs closer together and push unrelated pairs apart. The result is a space where semantic relationships become spatial ones, and 'is this similar to that' becomes a question you can answer with geometry instead of keyword matching.