You just want to make sure you trigger a snapshot, after you index with automatic embeddings before switching to a CPU node, so that the generated embeddings are persisted, instead of the original documents being replayed from the raft logs on restart and auto-embedding running again on the raw documents again