Skip to main content
Natural Language Processing

Reranking

A post-processing step that re-evaluates initial search results to reorder them by higher relevance

#Reranking#Re-ranking#Search post-processing#Reordering#RAG

What is Reranking?

Reranking is the stage where document candidates retrieved by a first-pass search are re-scored and reordered.

Typically, vector search or hybrid search pulls the Top-N candidates quickly, and then a more sophisticated model or rule set reshuffles them into a final ranked list.

Why Is It Needed?

First-pass retrieval is optimized for speed, so irrelevant documents may sneak into the top results. Applying a reranker improves the quality of the context fed into answer generation, which in turn lifts RAG answer accuracy.

Practical Application Points

  • First-pass retrieval: fast candidate recall (e.g., Top-20)
  • Second-pass reranking: accuracy-first reordering (e.g., pick Top-5)
  • Monitor metrics together: Precision@K, ground-truth inclusion rate, and response latency

Related terms