What is Reranking?
Reranking is a second-stage ranking operation that scores a query together with already retrieved candidates and reorders those candidates for a specific task. It usually follows lexical, dense, or hybrid retrieval because pairwise scoring is more expensive than independently indexing documents.
Quick Facts
| Full Name | Reranking in Information Retrieval and RAG |
|---|---|
| Created | Widely adopted as a core solution for retrieval bottlenecks during the industrialization of RAG technologies in 2023-2024. |
| Specification | Official Specification |
How It Works
Reranking improves the ordering of a candidate set; it does not create candidates that retrieval never returned, prove that a passage is sufficient evidence, or authorize a document for the current user. A cross-encoder is a common reranker: it reads a query and a candidate together, captures interactions that independent embeddings may miss, and returns a model-specific score. Other rerankers may use late interaction, learned features, or task-specific rules, so implementation should not be inferred from the label alone.
Cross-encoder scores are ranking signals, not portable probabilities. Some models expose logits rather than values between zero and one, and score distributions change with the model revision, language, query shape, document length, and candidate source. Calibrate thresholds and any abstention rule on labeled workload slices instead of copying a cutoff from a vendor example.
A production evaluation holds the corpus, permissions, queries, source revisions, and candidate-generation policy constant while comparing a baseline and a reranked variant. Measure candidate Recall@K before reranking, then nDCG, MRR, Precision@K, answer-evidence support, p50 and p95 latency, cost, and failure slices. Version the retriever, reranker, prompt or context assembler, policy, and evaluation set together so a ranking change remains explainable.
Key Characteristics
- Second-stage operation that reorders an existing candidate set rather than searching an entire corpus by default
- Often uses a cross-encoder that jointly processes a query and one candidate, without producing reusable document embeddings
- Bounded by candidate recall: a relevant source absent from the candidate set cannot be promoted
- Uses model-specific scores that require workload-specific calibration before they drive thresholds or automation
- Introduces latency and cost proportional to candidate count and input length, so depth must be evaluated with quality
- Requires authorization and source-lifecycle filtering before private content reaches the reranker
Common Use Cases
- Reordering hybrid or dense retrieval candidates before assembling evidence for a RAG answer
- Ranking support articles after lexical and semantic candidate generation
- Selecting the most relevant passages from a small, already authorized document scope
- Comparing retrieval variants through offline relevance and answer-evidence evaluation
- Applying a task-specific ranking model after filters for tenant, language, freshness, and document type
Example
Loading code...Frequently Asked Questions
How is reranking different from retrieval?
Retrieval generates a candidate set from a corpus, often using lexical, dense, or hybrid indexes. Reranking compares the query with each candidate more closely and changes their order. A reranker cannot promote a relevant document that candidate generation excluded.
Is every reranker a cross-encoder?
No. Cross-encoders are a common pairwise reranking architecture, but late-interaction, learned-feature, rule-based, and other task-specific rankers also exist. The interface should record the ranking model and score semantics rather than assume every reranker has the same behavior.
Can I use a fixed reranker score threshold?
Not safely across models or workloads. Some models emit logits, while others use different scales. Calibrate any keep, reject, or abstain threshold against labeled queries, languages, document types, and hard negatives for the deployed model revision.
How many candidates should a reranker score?
There is no universal Top-K. Increase candidate depth until relevant evidence reaches the reranker often enough, then measure whether ranking gains justify p95 latency and cost. Evaluate candidate recall and final ranking together under the same filters and corpus.
Does reranking prevent RAG hallucinations?
No. Better ordering can improve the evidence presented to a model, but it does not prove source authority, completeness, freshness, permission, or claim support. Combine reranking with retrieval evaluation, evidence sufficiency checks, citation validation, and an explicit no-answer policy.