The prediction happens when it is uncompressed right?

LLM embeddings are compressed training data.

To decompress that is to make a prediction (in this case to convert the embedding into readable text)