Reranking is a second pass over search results. A fast first stage retrieves a handful of candidate passages; a slower, sharper model then re-scores each one against the query and re-orders them, so the passage that genuinely answers the question rises to the top.
Why a second pass
First-pass retrieval is built for speed. Semantic search compares a single vector embedding of the query against pre-computed embeddings of every passage, which scales to millions of chunks but compresses each one into one fixed vector. That compression is lossy: the top result by vector similarity is usually about the right topic, but not always the passage that most directly supports the answer.
A reranker spends more compute on far fewer candidates. Instead of comparing two separate embeddings, it reads the query and a passage together and judges how well that specific passage answers that specific query. It is too slow to run across a whole corpus, but ideal for re-scoring the twenty or so hits the first pass already narrowed down.
How it sharpens a citation
This matters most for where a citation lands. In retrieval-augmented generation, the model writes its answer from the passages handed to it - so the order of those passages shapes which sentence a claim is grounded in. If the first pass ranks a loosely related paragraph above the one that actually states the fact, the answer may cite the wrong line.
Reranking corrects that ordering before the model ever sees it. By promoting the passage that truly contains the supporting sentence, it makes the resulting citation point at the precise text a reader should check, rather than something merely nearby.
Where it fits in Sidenote
When you ask Sidenote about a document, candidate passages are reranked before the model answers, which is how a citation can scroll to the exact sentence behind a claim rather than just the right page.
FAQ
Why not run the reranker on every passage?
Cost and speed. A reranker reads the query and a passage together, so scoring a million chunks that way would take minutes per question. Embedding search is cheap enough to scan everything; the reranker's job is only to sort the shortlist.
Does reranking make answers more accurate?
It makes citations more precise, which is most of what accuracy means in document Q&A. The model can only ground an answer in the passages it is handed, so promoting the passage that truly contains the fact directly improves which sentence gets cited.