GLOSSARY
Reranking
A second-pass model that re-sorts search results by true relevance — the quiet quality boost behind serious RAG and enterprise search.
Vector search is fast but shallow: it compares compressed geometry, so roughly-relevant items jostle with truly-relevant ones. A reranker reads each candidate fully — query and document side by side — and scores exact relevance. The pipeline becomes: retrieve 100 candidates fast, rerank to pick the best 5, answer from those 5.
Cohere built a business on exactly this component, and every production RAG stack eventually asks itself whether to add one. The cost is an extra model call; the benefit is that the language model answers from the right passages instead of plausible-but-wrong ones — which, given how grounding failures compound, is usually worth it.