A search system can use a fast first pass to select candidates, then run a more detailed model on the short list. For example, the second stage can compare each document with the full question.
Reranking adds processing time and can only reorder candidates that reached it. A relevant document missed by the first pass cannot be recovered by reordering the current list.
ELI5
Imagine pulling ten likely books off a shelf by reading their titles. You then skim their contents and move the most useful books to the top.
The second look is like reranking. You spend more time on a small group instead of reading every book in the library.
Can reranking recover a result missing from the candidate list?
It can only reorder the candidates it receives.
When does reranking help?
It can improve ordering when the first pass finds relevant candidates but ranks some of them poorly.

