Reranking / Re-ranking
Definition
Reranking is a second-stage ranking process that reorders an initial set of retrieved candidates using a different or more computationally expensive relevance model.
Why it matters
Initial retrieval often prioritises speed and broad candidate coverage. Reranking allows the system to apply deeper query-document comparison to a smaller candidate set before selection or answer generation.
How it works
Initial candidate set
- A retriever or fusion stage returns a manageable set of candidate documents, passages, products or entities.
Secondary scoring
- A reranker evaluates those candidates using additional features or a model that can perform richer query-candidate comparison.
- Cross-encoders and other learned ranking models are common examples.
Reordered output
- Candidates are sorted according to the reranking scores.
- Only the highest-ranked items may proceed to the next stage.
How Netsleek uses the term
Netsleek uses Reranking to distinguish initial retrieval from later candidate ordering. This is important because being retrieved does not mean a brand, passage or source will remain prominent after a more selective ranking stage.
Comparisons
- Reranking vs Retrieval: Retrieval identifies candidates. Reranking reorders candidates that have already been retrieved.
- Reranking vs Ranking Functions: Ranking Functions is the broader category. Reranking specifically refers to a later ranking pass over an existing candidate set.
- Reranking vs Selection: Reranking orders candidates. Selection determines which candidates are ultimately used, cited or recommended.
How Reranking Connects Across AI Systems
Reranking / Re-ranking does not operate in isolation. It connects to other established concepts across the AI search and retrieval process:
- Follows candidate generation: Candidate Generation supplies the items available to the reranker.
- Can follow fusion: RRF or other fusion methods can produce a candidate list that is subsequently reranked.
- Feeds selection: Higher-ranked candidates may be more likely to enter context or later selection stages.
- Affects exposure: A candidate retrieved initially can lose prominence after reranking.
Related glossary concepts
- Candidate Generation
- Ranking Functions
- Reciprocal Rank Fusion (RRF)
- Hybrid Search
- Selection Architecture
- AI Recommendation Layer
- Recall (Information Retrieval)
Common misinterpretations
- Reranking does not retrieve entirely new candidates unless the system runs another retrieval stage.
- A high reranking position does not guarantee final recommendation.
- Not every search system uses a distinct reranking stage.
Summary
Reranking applies a second ranking stage to an already retrieved candidate set. It helps modern search systems combine fast broad retrieval with more selective downstream relevance evaluation.