How would you solve inference for things like face detection or object detection? I guess you could run it on the client while uploading but that would make updating models awkward (download every file, run inference, send requests to server to update index).