Reranking
A second scoring pass that reorders retrieved candidates by relevance and quality before the best few are passed to the model.
In depth
What it really means
Retrieval is fast and approximate. It pulls maybe fifty candidates. Reranking is slow and precise, scoring each candidate against the query properly, then keeping the top three to five. Most of what gets retrieved dies here.
This is the stage where quality starts to count rather than just topical match. A reranker can tell that your passage answers the question directly while a competitor’s merely mentions the same terms. Being retrieved is not the win. Surviving the rerank is.
How it works
- Initial retrieval returns a large candidate set, fast and approximate.
- A cross-encoder scores each candidate against the query in detail.
- Candidates are reordered by that score.
- The top few are passed into the context window. The rest are dropped.
Pros & cons
Pros
- Directly answering the question beats keyword overlap at this stage.
- It filters out topically close but unhelpful content.
- Well-written, complete passages win over padded ones.
Cons
- Only a handful survive, so the funnel narrows sharply.
- You cannot see whether you were retrieved and dropped, or never retrieved.
- Reranker behavior varies by system and is not documented.
Common mistakes
- Optimizing to be retrieved and stopping there.
- Passages that mention the topic without answering the question.
- Burying the answer under context, so the top of the passage scores poorly.
- Padding sections, which dilutes the relevance score.
Best practices
- Answer the question in the first sentence of each section.
- Make the passage complete enough to stand alone as the answer.
- Cut padding, since density of relevant content is what scores.
- Match the question’s phrasing in the heading directly above the answer.
- Be specific. Concrete answers outscore general ones on the same topic.
FAQs
What is reranking?
A second, more precise scoring pass over retrieved candidates, reordering them by relevance and quality before the top few reach the model.
Why does reranking matter for content?
It is where most retrieved content gets eliminated. Being pulled into the candidate set is not enough. The passage has to actually answer the question better than the others.
How do I survive reranking?
Answer the question directly in the first sentence, keep the passage complete and self-contained, and cut padding that dilutes relevance.
Can I tell if I was reranked out?
No. There is no visibility into the pipeline. You only see whether you appear in the final answer.
Keep reading
Related on LymLyt
Beyond LymLyt
Further reading
Want this working on your site?
We build the content behind the term, ranked in search and cited by AI.
Book a 30-min call →