Embeddings represent words or text as dense numeric vectors. A token is a discrete, symbolic unit; an embedding is the continuous vector the model maps that token to — the model first tokenizes text, then looks up each token's embedding.

Similar meanings tend to have nearby vectors, enabling semantic similarity, clustering, retrieval, and downstream model features. Concretely, "king" becomes a token ID, and that ID maps to a high-dimensional vector like [0.21, -0.45, 0.87, ...] — hundreds or thousands of dimensions, learned so that related concepts end up geometrically close in that space. The distinction matters because tokens alone can't be reasoned about mathematically — they're just symbols — while embeddings give the neural network something continuous to compute over, which is what lets it capture semantic relationships instead of just matching strings.