How would you answer an interview scenario involving Vector search and similarity?

Explain how you would tune top-k and similarity thresholds for an FAQ assistant that sometimes retrieves irrelevant but semantically related documents.