179 questions
No questions match those filters.
← All questions
Retrieval & EmbeddingsVery Hard· HNSW· LSH· System Design
We’re replacing HNSW with LSH. The corpus takes 2M inse...
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 plansLSH allows for exact inserts and deletes because buckets are computed from the document alone, whereas HNSW and IVF are fit to the corpus structure. Before switching, compare the memory overhead (LSH is roughly 14x larger for equal recall) and candidate list size. If the rebuild window is the primary issue, consider segment-based indexing or a hybrid approach using LSH for hot data and a graph index for the cold majority.