← All questions
Retrieval & EmbeddingsHard· LSH· Latency· Evaluation

Your LSH index passes the offline gate at 99% recall, b...

Recall and the number of candidates both rise with L. By increasing L to hit the recall target, the system is now reranking a massive percentage of the corpus on every query, causing the latency spike and memory exhaustion. To fix this, lengthen the hash code to restore selectivity, reduce L, and re-measure recall.

Code playground (Python)