Re-ranking
A second model re-orders the top 20 results by relevance, then you keep the best few.
Last updated
Analogy
A quick first-round screen, then a proper interview for the shortlist.
In an agent
Better precision for a small extra cost, so the agent reads fewer useless chunks.
Practice
Flashcards
- What is re-ranking?A second pass that re-orders retrieved results by relevance.
Interview questions
- Your RAG gives wrong answers. How do you debug it?
- Check the trace: were the right chunks retrieved?
- If not, fix retrieval: chunking, hybrid search, re-ranking, metadata filters.
- If they were, fix the prompt or the model.
- Then add the failing case to the eval set so it stays fixed.
Questions people ask
What is re-ranking in RAG?
Re-ranking is a second pass over search results. You retrieve a wider set, for example the top 20 chunks, and a re-ranking model scores each one against the question more carefully. You keep only the best few. It improves precision for a small extra cost per query.
Is re-ranking worth the extra cost?
Often, yes, when answers suffer from irrelevant chunks. The first vector search is fast but rough; the re-ranker reads each candidate against the question. Sending fewer, better chunks to the model can save more tokens than the re-ranker costs. Measure it with retrieval evals before and after.