A few ideas - set backup models to query from - gracefully fail by generating an alternative from a local model