Cross-Encoders vs Bi-Encoders Quiz

Answers and explanations appear as you take the quiz.

1. What is the defining architectural difference between a cross-encoder and a bi-encoder?
   • A cross-encoder processes both texts together so tokens can attend across them, while a bi-encoder encodes each text separately
   • A bi-encoder uses attention and a cross-encoder does not use attention
   • A cross-encoder is smaller and faster than a bi-encoder
   • A bi-encoder can only handle one text at a time
2. Why do bi-encoders scale to millions of documents efficiently?
   • They use less training data
   • Document embeddings are query-independent and can be precomputed and stored
   • They never use transformers
   • They skip the slow tokenization step
3. In a typical production search pipeline, how are the two architectures combined?
   • Cross-encoder retrieves candidates, then a bi-encoder reranks them
   • Both run in parallel and votes are averaged
   • A bi-encoder retrieves many candidates, then a cross-encoder reranks the top ones
   • Only one is ever used, never both
4. How does a late-interaction model such as ColBERT combine token-level query and document representations at scoring time?
   • It averages both texts into a single vector before any interaction
   • It concatenates raw audio waveforms and applies a vocoder
   • It ranks documents only by their character count
   • It compares query-token vectors with document-token vectors using a MaxSim-style aggregation
5. What does a bi-encoder typically output for a single input text?
   • A single fixed-length embedding vector
   • A relevance score between two texts
   • A ranked list of documents
   • A probability distribution over the vocabulary
6. After jointly processing a query and candidate, what does a cross-encoder produce?
   • A reusable document vector computed independently of the query
   • A relevance score for that specific query-candidate pair
   • A separate token vector for every item in the full corpus
   • A guarantee that the top result is correct
7. What kind of representation does a late-interaction model such as ColBERT retain for a text?
   • Only one global label for the whole corpus
   • A decoded audio waveform
   • A vector representation for each token
   • A complete copy of the cross-encoder’s pairwise scores
8. What similarity function do Sentence Transformers examples commonly use for comparing normalized bi-encoder embeddings?
   • Edit distance over raw audio samples
   • A pixel-wise image loss
   • A speech-recognition word error rate
   • Cosine similarity