Maybe, except from my understanding an embedding vector tends to be much larger than the source token (due to the high dimensionality of the embedding space).

Maybe, except from my understanding an embedding vector tends to be much larger than the source token (due to the high dimensionality of the embedding space). So it's almost like a reverse compression in a way. That said I know vector DBs have much more efficient ways of storing those vector embedding.