Reranking
Reranking is a second pass in retrieval where an initial set of candidate passages is reordered by a more precise relevance model, deciding which few actually make it into the AI's answer.
Reranking is the cut from "retrieved" to "actually used." A first, fast search pulls a broad set of candidate passages; a slower, sharper model then scores those candidates for true relevance and reorders them, and only the top few survive into the context the AI writes its answer from. Being retrieved isn't enough — you have to win the rerank.
This is where the quality of your passage earns its keep. Reranking models reward passages that directly and completely answer the specific question, so a tight, self-contained, on-point paragraph — the extractability pillar — outranks one that's merely topically related. Signals of trust and authority can also tilt which of several good candidates gets promoted when relevance is close.
Example. A query might retrieve 50 candidate passages, but the reranker keeps only the best 5 for the model to read. If your paragraph is relevant but buried in hedging and tangents, it gets retrieved and then dropped at the rerank — present in the pool, absent from the answer.
Relevant pillars
Related terms
- Vector SearchVector search is a retrieval method that finds passages by meaning rather than keywords, comparing the numeric embedding of a query against the embeddings of indexed content to surface the closest matches.
- Passage RetrievalPassage retrieval is the practice of finding and returning specific relevant passages from within documents, rather than whole pages, which is why AI engines cite paragraphs instead of articles.
- Position BiasPosition bias is the tendency of retrieval and language models to weight content near the start of a page or passage more heavily, making where you place an answer matter as much as the answer itself.