CORE GUIDE
Reranking retrieved candidates
Reranking is a second selection stage that spends more computation on a smaller candidate set to order evidence by task relevance after broad retrieval maximizes recall.
Mental model
Retrieval and reranking solve different objectives: first retrieve enough plausible candidates cheaply, then apply richer query-document reasoning to decide which few pieces deserve scarce context budget.
Why it matters
Vector or lexical retrieval often finds the right evidence but does not place it high enough for the generator to use. Increasing top-k indefinitely wastes context and can add distracting passages. A reranker can use richer features or a cross-encoder-like comparison to improve ordering, but it still cannot repair candidates that the first stage never retrieved.
01
Separate recall from precision
Evaluate first-stage recall at a candidate depth that captures relevant evidence, then evaluate reranking quality on those candidates. Feed the reranker the query and candidate text plus only stable metadata it can interpret. Measure important slices such as exact identifiers, recency and policy questions, and keep authority filters independent when correctness depends on source hierarchy.
02
Example: correct policy is retrieved at rank twelve
A hybrid retriever returns twenty candidates and the current refund policy appears at rank twelve behind broadly similar FAQ pages. A reranker recognizes the query's policy-specific wording and moves the controlling document near the top. If the current policy had not appeared in the candidate set at all, reranking could not recover it.
Common failure modes
- Blaming reranking for failures caused by missing first-stage recall.
- Increasing candidate depth without measuring latency and distractor cost.
- Letting relevance ranking override explicit source-authority rules.
Engineering heuristics
- Measure candidate recall and final ranking separately.
- Inspect rank changes on labeled query slices.
- Apply authority and hard metadata constraints independently from learned relevance.
Takeaways
- 01Reranking improves order, not missing recall.
- 02Rich ranking belongs after broad candidate generation.
- 03Retrieval evaluation should identify which stage caused the failure.
Reading evidence
This records actions you actually took; it does not claim mastery, proficiency, or certification.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.