Comment by alanwli

Out of curiosity, how is the 92% recall calculated? For a given query, is the recall compared to the true topk of all 100B vectors vs. recall at each of N shards compared to the topk of each respective shard?