179 questions
No questions match those filters.
← All questions
Retrieval & EmbeddingsMedium· Cross-encoders· Inference Latency· Reranking
Why is a cross-encoder not used for retrieval, and how...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansA strong answer provides the arithmetic: 2P (Lq + Lp) FLOPs per pair, which for millions of passages is impossible within a 50ms budget. The bi-encoder is efficient because the passage tower is query-independent and runs offline. The cross-encoder is not discarded; it is moved to a later stage (e.g., depth 100) where it can provide high-precision ranking within a manageable latency window.