Treat missing evidence as an ordinary product outcome. SearchCandy exposes not_covered so your application can decide what to do next.
r = sc.query(question)
if r.not_covered:
message = "No supporting passage returned. Add a relevant document or ask for help."
else:
message = "\n".join(q.text for q in r.quotes)Do not convert a missing-evidence result into an invented answer. A useful interface offers a next action: refine the question, add relevant material, or ask a person.
A retrieval system can miss relevant material. Evaluate both correct refusals and cases where evidence exists but is not returned. The documented result of 55 out of 55 absent-entity questions refused describes that evaluation, not every possible question.
An indexing pause, invalid credential, or unavailable service is not the same as missing evidence. Handle the documented error response separately and show an appropriate retry or access message. Do not label an API failure as “not in your documents.”
High quality among answered questions can coexist with missed questions. Report the number of questions tested, answered, correctly refused, incorrectly refused, and failed by the service. Keep ungraded and excluded questions visible in your accounting.
Next: Plan a pilot evaluation →