Reranking applies a second, more computationally intensive model to score the top results from an initial retrieval pass. The initial pass optimizes for recall across a large index, while reranking optimizes precision among the candidates.
It is a standard component of production retrieval systems.
Alternative Names:
Rerank, Result Reranking
Why it Matters?
Reranking addresses the practical constraint that only a limited number of passages fit in a model's context. If the genuinely relevant passage ranks twentieth and only the top five are supplied, the answer will be wrong regardless of model quality. In matters with large document sets, reranking quality is frequently the difference between an accurate answer and a confident wrong one.
Frequently Confused with
Related terms
Frequently asked questions
Why is reranking necessary?
How does it differ from initial retrieval?
