Reranking

Reranking is a second, slower pass over search results: a cross-encoder model reads each candidate together with the query and rescores it for true relevance, so the top results are the actually-right ones, not just the roughly-similar ones.

Data & retrievalre-rankingresult rescoring
§ 01DefinitionReranking

No jargon, no math

In plain words, what is it?

Vector search is fast because it compresses the query and the documents separately: fast, but shallow. Reranking adds a second read. First pass: an embedding search pulls the top 50 candidates in a few milliseconds, tuned for recall. Second pass: a cross-encoder reads the query and each candidate together, pair by pair, and assigns a precise relevance score; the list is reordered and only the top 5 reach the model that writes the answer. The cost: joint reading is slow, which is why you reserve it for a shortlist. The limit: if the right document was not in the first-pass 50, reranking never sees it, so the first pass is tuned to avoid losing it.

§ 02How it worksFast 50 first, slow precise top 5 after
  1. 01

    First pass. A fast embedding search pulls the top 50 candidates.

  2. 02

    Joint reading. The cross-encoder reads the query and each candidate together, pair by pair.

  3. 03

    Rescoring. Each pair gets a precise relevance score and the list is reordered.

  4. 04

    Hand-off. Only the top 5 reach the model, quick to read and cite.

§ 04

Want this running in your stack?

We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.