Reranking is the stage in a search or retrieval pipeline where an initial set of candidates, found quickly by keyword or vector search, is reordered by a slower and more accurate model. The reranker reads the query and each candidate together and scores how well the candidate answers the query. In AI search, this stage helps decide which handful of passages make it into the model's context window.
Why two stages
First-stage retrieval has to search millions or billions of documents in milliseconds, so it uses fast methods that compare precomputed representations. Those methods are good at finding plausible candidates but less precise about which candidate best answers the question. A reranker reads the query and passage side by side, which is more expensive but more accurate. Running it only on the top candidates keeps the cost manageable.
- Stage 1: retrieve a broad shortlist quickly
- Stage 2: rerank the shortlist with a stronger model
- Stage 3: pass the top passages to the language model
What a reranker tends to reward
Rerankers are trained to judge whether a passage answers a query. That tends to favor passages that address the question directly and specifically. A passage that circles a topic without stating an answer may survive the first stage on topical similarity and then lose at reranking to a passage that actually answers.
The exact models and signals used by commercial AI search products are not public, so treat this as a description of the general technique.
Example
Two passages about the same feature
A user asks an assistant whether Fernhollow, a fictional HR platform, supports SCIM provisioning. Two Fernhollow passages are retrieved. One is a marketing paragraph about 'enterprise-grade identity integrations'. The other says 'Fernhollow supports SCIM 2.0 provisioning with Okta and Microsoft Entra ID on the Business plan.' A reranker is far more likely to place the second passage on top, because it answers the question.
Common mistakes
Optimizing only for topical relevance
Being about the topic gets you into the shortlist. Answering the question directly is what a precise ranking step looks for.
Hedging every statement
Vague passages are hard for any model to score as a clear answer. State facts plainly where you can support them.
Questions
Web search engines also use multiple ranking stages, but 'reranking' in AI discussions usually refers to the passage scoring step inside a retrieval pipeline.
Not for commercial engines. You can approximate with open reranking models on your own content to compare passages against target questions.
