Reranking
Which retrieved passages deserve attention first?
A reranker reorders an existing candidate set for a query. Cross-encoder rerankers examine query and document together, allowing closer relevance judgments than independently encoded vectors alone. Reranking can improve the material selected for context, but it cannot add missing evidence or verify the truth of a source.
Query“replace Model X battery”
1A general guide to batteriesBroad background
2Model X: replacing the batteryMatches the model and the task
3Model X repair manualRelevant, with a wider scope
Retrieval gathers plausible candidates. The first result is not always the best evidence for this task.
Illustrative ordering, not a model benchmark.
What to understand
- Retrieve a candidate pool first, then rerank and select the material that fits the answer's context budget.
- A relevance score describes model output for a query-document pair. It is not a probability that the document is true.
- Measure the relevance gain against added latency and cost, and check whether long inputs are truncated.
Go to the source
Primary documentation for the ideas in this explainer.
Follow the next part of the system.
Evidence & citations