Retrieval Reranking Quiz

Answers and explanations appear as you take the quiz.

1. What is the role of a reranker in a retrieval pipeline?
   • To re-score a small candidate set so the most relevant results rise to the top
   • To scan the entire corpus from scratch
   • To generate brand-new documents
   • To compress the search index
2. What distinguishes a cross-encoder from a bi-encoder?
   • A cross-encoder embeds query and document separately
   • A cross-encoder feeds query and document together so attention compares them jointly
   • A cross-encoder cannot use transformers
   • A cross-encoder only works on images
3. Why are cross-encoders used for reranking rather than first-stage retrieval?
   • They are inaccurate
   • They only handle one document ever
   • They are too slow to run over an entire corpus because document vectors can't be precomputed
   • They require no computation
4. What does the first-stage retriever typically optimize for?
   • Precision at rank 1 only
   • Generating natural language answers
   • Reordering exactly five documents
   • Recall and speed across the whole corpus
5. How can a reranker affect a RAG candidate set, and what can it not guarantee?
   • It can reorder retrieved candidates, but cannot add a missing passage or verify generated claims
   • It guarantees the most relevant source is present and every generated claim is correct
   • It replaces the first-stage retriever and searches the entire corpus
   • It makes retrieval precision and recall perfect
6. Which of these is an emerging reranking approach?
   • Deleting the retriever and guessing answers
   • Using LLMs as listwise rerankers that reason over the whole candidate set
   • Reverting to manual keyword tagging only
   • Replacing transformers with random scoring
7. Which metric can assess whether the first-stage retriever includes relevant passages before reranking?
   • Image resolution of the generated answer
   • BLEU score of an unrelated translation
   • Recall of relevant candidates in the retrieved set
   • Number of model parameters in the reranker
8. Which metric did the Passage Re-ranking with BERT paper report a 27% relative improvement on for MS MARCO?
   • BLEU-4
   • ImageNet top-1 accuracy
   • Mean absolute error
   • MRR@10