This is something I want to make possible! There was an early experiment in https://github.com/mcclure/cuervo/ which helped identify some ways that our embedding APIs would need to change to support that use case.