ScienceTowards Data Scienceabout 2 hours ago
How Does a RAG Reranker Really Work?
1 min read
A RAG reranker is a cross-encoder that scores query-passage pairs jointly, but its dominant learned signal is keyword co-occurrence, not answering. It runs 30 to 100 times slower per query than an embedder. The mechanism transfers only when corpus vocabulary resembles MS MARCO training data.
Level
Hype check
Tap to vote and see what everyone thinks.
Summary by ByteBrief
