← All questions
RAGHard· Retrieval· Metrics· Optimization

Your similarity gate refuses 11.5% of questions that ha...

Lowering the cutoff is a mistake because it trades away the unanswerable queries you were successfully catching. The issue is that the softmax objective does not constrain the score level, meaning no global cutoff is optimal. The fix is to change the measurement method: implement an entailment pass over the candidate answer and top chunks, which provides higher precision and better alignment with the goal of verifying support.

Code playground (Python)