Reranking
A fast search can find the right passage but rank it 9th. A reranker reads the question and each candidate passage together and scores them carefully. It moves the best ones to the top. So the LLM gets a few great chunks instead of many average ones.
New words
- Reranker
- A slower, more careful model that re-orders search results.
- Recall
- Did the candidate list include the good passages at all?
- Precision
- How many of the passages you keep are actually relevant.
- Bi-encoder
- Turns question and passage into vectors separately. Fast.
- Cross-encoder
- Reads question and passage together and gives one score. Accurate.
- Candidate
- A passage from the first search that might answer the question.
What you will learn
- The problem: the right answer is buried
- How it works: two-stage retrieval
- Try it: rerank the candidates
- Trade-offs and failure modes
- Production practices
- In code
- In a real system
- Interview tip
- Key takeaways
10 quiz questions at the end.
The full lesson, with interactive demos and the quiz, opens after you sign in with Google. It is free.