I ported over the test expressions is using here. Basically we just apply a marker interface if scores are needed and then we can check with instanceof. I guess alternatively it could be an annotation or something like that, but this seems simple enough?

Separately we should followup with the special score handling. Maybe the scoreAccessor should not just be shoved into the input map like we do, but instead just passed in as a normal parameter (null if not needed or n/a)? This would avoid unnecessary per-document hashing.

Interface is fine, annotation is IMHO too complicated, but can be done without much changes.

LGTM.