What is reranking?
A second-stage retrieval step that reorders candidate chunks using a stronger relevance signal.
Interview Questions: Reranking and Retrieval Optimization
When does reranking help?
When the right evidence is retrieved but ranked too low for final context.
What can reranking not fix?
It cannot recover evidence missing from the candidate pool.
What is retrieve many, return few?
Retrieve a broad candidate pool, rerank it, then send a smaller high-quality context set.
Bi-encoder vs cross-encoder?
Bi-encoders are fast for search; cross-encoders read query and candidate together for slower but often better relevance scoring.
What metrics matter?
Candidate Recall@k, MRR, gold final rank, citation support, latency, timeout rate, and cost.
How do you manage latency?
Limit candidate pool, route reranking to high-value cases, use timeouts, and fallback to first-stage results.
Why dedupe after reranking?
To prevent repeated chunks from crowding out diverse evidence.
How do you release reranking?
Shadow test, compare before/after metrics, monitor latency, then roll out gradually.
What is the senior answer?
Reranking is useful when measured candidate recall is good but final ordering is weak.