ColBERT Late Interaction Retrieval Quiz

Answers and explanations appear as you take the quiz.

1. How does ColBERT represent a query or document?
   • As a single averaged embedding vector
   • As raw keyword counts only
   • As many token-level contextual embedding vectors
   • As an image
2. What is the scoring function ColBERT uses called?
   • MaxSim
   • Cosine pooling
   • Softmax cross-entropy
   • Beam search
3. What query-time work does ColBERT avoid by precomputing document token vectors?
   • All query encoding and all index access
   • Every comparison between a query token and document token
   • The need to store document representations
   • Joint transformer encoding of every query-document pair, while query encoding and similarity search still run
4. What does the 'late' in 'late interaction' refer to?
   • The model is trained late in the pipeline
   • The interaction between query and document tokens is deferred to the scoring step after independent encoding
   • Queries arrive late at night
   • Documents are added after deployment
5. What problem with single-vector dense retrieval does ColBERT address?
   • It is too slow to index
   • It cannot run on GPUs
   • It requires labeled data for every query
   • Compressing a whole passage into one vector loses fine-grained detail
6. What did ColBERTv2 notably improve over the original?
   • It removed the need for any embeddings
   • It switched from BERT to a rule-based system
   • It added residual compression to greatly shrink the index size
   • It eliminated the MaxSim step
7. A two-token ColBERT query has best document-token similarities of 0.8 and 0.3. What MaxSim score results?
   • 0.8, retaining only the single largest value across the whole query
   • 1.1, the sum of the two per-query-token maxima
   • 0.55, averaging instead of summing the maxima
   • 0.24, multiplying the two maxima
8. Why can a ColBERT document index require more storage than a single-vector dense index?
   • It retains multiple token-level representations for each document
   • It never precomputes any document features
   • It must store all transformer gradients for each document
   • It removes every document token before indexing