Recall@K is the primary metric for measuring the accuracy of approximate-nearest-neighbour search. For a given query, it asks: of the K vectors that are truly nearest, how many did the approximate search actually return in its top K results? A Recall@10 of 0.95 means the search found 95% of the genuine ten nearest neighbours.
This metric exists because approximate search trades some accuracy for speed, and Recall@K quantifies exactly how much accuracy is retained. A value of 1.0 means the approximate results perfectly match the exact answer; lower values mean some true neighbours were missed. It is computed by comparing approximate results against ground truth produced by an exact, brute-force search over a test set.
Recall@K always exists in tension with query speed. Any ANN index can achieve higher recall by searching more thoroughly, at the cost of latency, and the tuning parameters of indexes like HNSW and IVF exist precisely to set this balance. Benchmarking a vector database means plotting Recall@K against queries per second, since the meaningful question is not recall alone but how much recall you get at a given speed.