Definition
A secondary processing stage in a RAG pipeline that re-orders a subset of candidate documents retrieved by an initial search using a more computationally expensive cross-encoder model. It improves retrieval precision by analyzing the direct semantic interaction between the query and each document, trading increased latency for higher relevance accuracy.
Retrieval (Bi-Encoders) is about finding candidates; Re-ranking (Cross-Encoders) is about scoring their exact relevance.
"An HR recruiter skim-reading 500 resumes to find 10 candidates (Retrieval), followed by a panel of experts conducting deep-dive interviews with those 10 to find the best fit (Re-ranking)."
- Cross-Encoder(Component)
- Bi-Encoder(Prerequisite)
- Precision at K(Evaluation Metric)
- Semantic Similarity(Underlying Concept)
Conceptual Overview
A secondary processing stage in a RAG pipeline that re-orders a subset of candidate documents retrieved by an initial search using a more computationally expensive cross-encoder model. It improves retrieval precision by analyzing the direct semantic interaction between the query and each document, trading increased latency for higher relevance accuracy.
Disambiguation
Retrieval (Bi-Encoders) is about finding candidates; Re-ranking (Cross-Encoders) is about scoring their exact relevance.
Visual Analog
An HR recruiter skim-reading 500 resumes to find 10 candidates (Retrieval), followed by a panel of experts conducting deep-dive interviews with those 10 to find the best fit (Re-ranking).