noun · AI technique
Reranking
also called re-ranking
A second opinion on search results. A closer look puts the most useful ones at the top before the AI reads them.
A second retrieval step that rescores and reorders an initial list of candidates with a more precise model.
A first search quickly returns a wide set of candidates. A slower, more careful model then scores each candidate against the query and reorders them, so the few passages that reach the context window are the best ones.
Real-life example
Your help bot searches and gets twenty passages, but the best one is ninth. Only the top three fit in the prompt, so the bot answers from the wrong ones. Adding a reranking step looks at all twenty more carefully and moves the best passage to the top.
Where the analogy breaks
-
Reranking only reorders what was found. If the first search missed the right passage, the reranker cannot bring it back.
-
It adds time and cost. Every candidate is scored by an extra model, so the list is kept short.
-
A relevance score is not truth. A passage can be the best match and still be wrong or out of date.