In simple words: Embeddings turn text into numbers that can reflect useful relationships.
Beyond token IDs
An embedding is a vector: a list of numbers learned so useful relationships can be represented. Similar words or contexts may have nearby representations. It is not a dictionary definition.
What embeddings help with
Embeddings can support search, classification, recommendations, and language models. Contextual models can give the same word different representations in different sentences. Older fixed word embeddings assign one representation to a word regardless of context.
See the flow
“Car” and “vehicle” may have related embeddings even though their letters differ. A simple word count would not capture that similarity directly.
Remember: Embeddings turn text into numbers that can reflect useful relationships.