The setup is pretty simple: take a sentence encoder and freeze it. Then train a small projection layer that maps those embeddings into "soft prompt" tokens for a language model. The LLM learns to reconstruct the original text from just those projected embeddings.
It's far from perfect. You probably can't reconstruct the exact meaning of the text, but you can get the general idea/vibe of the original input.
