We'll need this so we can use a pull- rather than push-based model for the reference extraction service.