What does re-ranking do in retrieval?
Why A is the answer
It reorders candidates by relevance to the query
Retrieval returns candidates quickly. Re-ranking scores them more closely against the query, so the best passages reach the model.
In detail
- Retrieval returns a set of candidates.
- Re-ranking scores them against the query and reorders them.
- The best passages then reach the model.
- Re-ranking adds some latency, so weigh it against your target.
Use case
An illustrative scenario, not a real incident.
A search returns 20 passages, and the best sits at 14. A re-ranker scores all 20 against the query and moves it to the top, so it reaches the model.
See it in different ways
flowchart LR A["Query"] --> B["Retrieve candidates"] B --> C["Rerank"] C --> D["Best passages"] D --> E["LLM"]
mindmap
root((RAG app))
Data
Delta tables
Chunks
Retrieval
Vector index
Reranking
Chain
MLflow
Agent Framework
Serving
Model Serving
Governance
Unity Catalogflowchart LR S0["Retrieve candidates"] --> S1["Score against the query"] --> S2["Reorder"] --> S3["Send the best to the LLM"]
Key point: Re-ranking reorders candidates by relevance.
Why the other options are wrong
- B. It writes new documents for the index
- It orders existing results, and writes nothing.
- C. It trains the embedding model
- It does not train the embedding model.
- D. It deletes old chunks from the table
- It does not change the source data.