What is reranking?

Two-stage retrieval

Searching a large collection has to be fast, so the first stage uses cheap methods: keyword scoring, or comparing a query embedding against stored embeddings in a vector database. These methods encode the query and each document separately, which is what makes them fast, but it also means they never look at the two side by side.

A reranker does look at them together. The most common kind, a cross-encoder, reads the query and one candidate passage as a single input and outputs a relevance score. That joint reading catches things separate encodings miss, like negation, who did what to whom, or whether the passage actually answers the question. Because it runs once per candidate, it is only practical on a short list, typically the top 20 to 200 results. Some systems use a language model as the reranker, which is more flexible and slower still.

An example

Query: when did the client reject the proposal. The first stage returns passages that mention clients and proposals, including one where the client accepted a proposal and one where the team rejected a client's idea. Both are close in embedding space. A reranker reading each passage with the query can tell that neither matches who rejected what, and moves the passage where the client turned the proposal down to the top.

What reranking can and can't do

• It can fix ordering, so the best candidate comes first instead of fifth.
• It can help decide when nothing is relevant, if its scores are used as a cutoff.
• It can't recover a passage the first stage never retrieved. Better recall has to come from the first stage, for example through hybrid search.
• It adds latency that grows with the number of candidates, so the list size is a trade-off between quality and speed.
• It is only as good as its training. A reranker trained on web questions may misjudge transcripts full of filler words or specialist jargon.

Why it matters for searching media

Transcripts are messy text: people repeat themselves, trail off and talk around a subject. That makes the gap between “mentions the topic” and “is the moment you meant” wider than in tidy documents, and ordering matters more. In MediaFind, exact and semantic search are run together and fused into one ranked list, and each result links to its exact timestamp so you can confirm the top hit with one click instead of trusting the ranking blindly. For the bigger picture, see semantic search and how we measure search quality.

Frequently asked questions

What is the difference between a retriever and a reranker?

A retriever quickly finds candidates from the whole collection. A reranker carefully reorders a short list of those candidates. Most good search systems use both.

Is reranking the same as re-running the search?

No. It only reorders results that were already found; it does not look at the rest of the collection.

Does reranking make search slower?

A little. The cost depends on how many candidates are reranked and how large the reranking model is.