What is it?
Reranking is a second-stage ranking step applied after an initial retrieval.
A first search quickly finds possible matches; a reranker examines those candidates more carefully and puts the most relevant first.
08 · Embeddings, Search & RAG · Topic 10
What is Reranking?
Reranking reorders retrieved results so the most useful candidates appear first.
Think of it like...
A librarian first pulls a stack of possibly relevant books, then sorts that stack to put the best matches on top.
How it works
1 · QueryRead the requestIdentify what the user needs.
2 · Initial retrievalFind candidatesSearch returns possible matches.
3 · RerankerScore relevanceCompare candidates with the query in more detail.
4 · Improved orderPut the best firstSort by the new relevance scores.
5 · Top resultsSend a smaller setPass the strongest results to RAG or the application.
Real-world examples
Everyday example · Shopping search
Input: waterproof walking shoes
Output: the most useful products first
AI example · RAG assistant
System: retrieves 20 candidates, then reranks them
Output: the top 3–5 passages sent to the model
What it is NOT
Reranking ≠ Retrieval
Reranking orders candidates already found; retrieval finds candidates.
Reranking ≠ Vector Search
Reranking is a second-stage step; vector search finds semantically similar vectors.
Reranking ≠ RAG
Reranking orders results; RAG uses retrieved information to help generate an answer.
Related concepts
Query → Initial Retrieval → Candidate Results → Reranking → Top Results → RAG / Application
Remember this
Reranking improves the order of already retrieved candidates before they are used.
Video unavailable
Video
No video is available for this topic yet.