What does “Reranker” mean?
A model that reorders an initial set of retrieved results to put the most relevant items at the top.
How Reranker Works
Retriever returns 50 candidates
Reranker scores each with the query
Reorder by score
Top 5 go to the model.
Practical Example
A search step quickly pulls 50 passages that might answer a question. A reranker reads the question with each passage and promotes the three that truly answer it to the top.
Why It Matters
Reranking improves the quality of what the model sees at modest cost. It adds a bit of delay.
Common Questions About Reranker
Q: Why not rerank everything?
Rerankers are slower per document, so they are applied to a short list from a fast first-stage search.
Q: What models are used?
Often cross-encoders, or sometimes a language model prompted to rate relevance.
Q: How much does it help?
Often a noticeable improvement in relevance, depending on the data.