Definition
Reranking is the second pass over a set of search results, where a slower, more accurate model reorders the shortlist before anything reaches the answer. Passage ranking is what produced that shortlist in the first place. Reranking decides which few of those passages the model actually reads.
At a glance
- CategoriesTechnical
- Related fieldsTechnical
- DifficultyAdvanced
At a glance
- Reranking runs on tens of candidates, not millions, which is what makes the slower model affordable.
- A reranker reads the question and the passage together instead of comparing stored representations.
- Only the top few survive, because they have to fit inside the model’s working space.
A second, slower look at the shortlist
Retrieval is a funnel with two scoring methods in it.
The first pass is fast and approximate. Chunks have an embedding computed in advance, and the system finds the nearest ones without reading anything. It is quick because the hard work happened before the question existed.
The second pass reads properly. A reranker takes each candidate alongside the question and scores how well that passage answers it, an approach from work like the BERT paper. It catches what the fast pass misses: a passage close in meaning that answers a different question.
Then the list gets cut, because whatever survives has to fit in the context window alongside the question and the instructions.
Retrieved is not the same as used
Your logs show an AI crawler fetching a page. The answer never mentions you. Both are usually true at once, and reranking is where they part company.
Being retrieved means clearing a rough similarity bar. Being reranked into the top few means a model read your passage against the question and judged it the better answer. The second bar is the one that decides citations.
The gap usually comes from one thing: a passage about the right topic that answers nothing. A section headed “Shipping” describing your logistics partnership is topically perfect for a delivery-time question and useless as an answer. The fast pass loves it. The reranker drops it for a page stating the number of days.
Earn the second read
Write passages that hold up when something reads them closely.
- Answer in the section’s first sentence. A reranker rewards the passage that resolves the question immediately.
- Match the question, not the topic. If buyers ask “how long does delivery take”, your heading and opening line answer that, rather than describing your shipping operation.
- Keep the passage complete alone. One that depends on the paragraph above reads as a partial answer and scores like one.
- Put the specific detail in. Numbers, conditions, names. Specificity separates a passage that answers from one that is merely relevant.
Frequently asked questions
-
Is reranking the same as passage ranking?
Two stages, same pipeline. Passage ranking is the wide first pass, scoring a large pool cheaply to produce a shortlist. Reranking is the narrow second pass that reads those finalists and reorders them.
-
Why not run the accurate model on everything?
Cost and latency. A reranker reads the question and each candidate together, so the work scales with candidate count instead of being precomputed. Running it across a whole index per question would be impossibly slow. The funnel exists so the expensive model sees a few dozen things.
-
Can I influence reranking directly?
Not as a lever of its own, and nothing here is configurable. You influence it the way you influence everything else in retrieval: passages that answer a specific question in their first line.
-
Does this explain why crawlers fetch my page but never cite it?
Often. A fetch means you passed a similarity check. A citation means a model read your passage against the question and preferred it. Crawled but never quoted usually means your sections cover subjects rather than answer questions.