ColBERT and Multi-Vector Retrieval Quiz

Answers and explanations appear as you take the quiz.

1. How does ColBERT represent a document?
   • As many per-token contextual vectors
   • As a single dense vector
   • As a sparse bag of keyword counts
   • As a raw text string only
2. What operation does ColBERT use to score query-document relevance?
   • Cosine of two single vectors
   • MaxSim: sum of each query token's best match to any document token
   • Cross-entropy over the whole vocabulary
   • BM25 term frequency
3. Why is ColBERT faster than a cross-encoder at scale?
   • It uses a smaller vocabulary
   • It ignores the query entirely
   • Document token embeddings are precomputed and indexed offline
   • It processes only the first word of each document
4. What problem with single-vector dense retrievers does ColBERT address?
   • They are too slow to index
   • They cannot use neural networks
   • They require no training data
   • They lose fine-grained, token-level detail by compressing to one vector
5. What key improvement did ColBERTv2 introduce?
   • Residual compression that greatly shrinks the index size
   • It removed BERT entirely
   • It switched to single-vector encoding
   • It only works on images
6. What does ColPali apply ColBERT-style late interaction to?
   • Audio waveforms
   • Image patches of document pages, like PDFs
   • Single keywords only
   • Database table rows
7. In a two-stage search system, what role can ColBERT play after a fast first-stage retriever?
   • Generate the source documents from scratch
   • Replace the query with a document title
   • Rerank the candidate passages using contextual token-level matching
   • Discard every candidate before scoring
8. What does ColBERTv2's residual compression store for a token vector?
   • Only the raw document text with no vector data
   • A full copy of the language model for each token
   • One shared vector for the entire document
   • A centroid code together with a compact residual representation