The numerical vector representing the embedding is only useful for finding documents that are similar to your search query.

Those documents are then injected into your prompt and sent to some kind of LLM completion system such as GPT.

So yes, you will be sending chunks of your actual notes over the wire.