Embedding model has to kept in RAM for inference.