Reranking Models Quiz

Answers and explanations appear as you take the quiz.

1. In a typical two-stage retrieval pipeline, what is the reranker's job?
   • To re-score and reorder a shortlist of candidates for relevance
   • To pull the initial candidates from millions of documents
   • To store documents in a database
   • To generate the final natural-language answer
2. Besides MS MARCO, on which dataset did Nogueira and Cho report strong results for their BERT passage reranker?
   • ImageNet
   • TREC-CAR
   • LibriSpeech
   • SQuAD only
3. Why can't a cross-encoder reranker be used to search millions of documents directly?
   • It only works on images
   • It cannot read text
   • Its scores must be computed fresh for each query-document pair, which is too slow at scale
   • It requires no computation at all
4. What is the main advantage of a bi-encoder over a cross-encoder?
   • It is more accurate at fine-grained relevance
   • It always reorders results perfectly
   • It needs no training data
   • Its document embeddings can be precomputed, making retrieval fast and scalable
5. How can a reranker affect context quality in a retrieval-augmented generation pipeline?
   • It can move more relevant retrieved candidates higher in the context shortlist, but does not guarantee correctness
   • It ensures every retrieved passage is factually correct
   • It creates missing documents in the corpus
   • It removes the need to evaluate retrieval
6. In listwise reranking, what does the model consider together?
   • Only one document with no query
   • A query and a set of candidate documents when deciding their ordering
   • Each document embedding before any candidate retrieval
   • A single relevance label for the full training corpus
7. What should a team verify before interpreting reranker scores as probabilities?
   • Whether every score is above zero
   • Whether the first-stage retriever was removed
   • Whether the scores have been calibrated and evaluated for that use
   • Whether the model has more parameters than the retriever
8. What limitation does a reranker inherit from the first-stage retrieval system?
   • It always returns a calibrated probability
   • It removes the need for retrieval
   • It guarantees a correct generated answer
   • It cannot promote a relevant document that was never included in the candidate set