ANN Search: Recall What Matters
Researchers propose a new metric, 1/Ratio@k, to evaluate Approximate Nearest Neighbor (ANN) search algorithms, arguing it better reflects retrieval quality than the traditional Recall@k. This new metric, which assesses the difference in distances between retrieved and true neighbors, is judge-free and computable from standard benchmark inputs. Experiments show that optimizing for 1/Ratio@k achieves operational quality at a lower computational cost and more accurately tracks downstream task performance, such as classification and retrieval-augmented generation, compared to Recall@k. AI
IMPACT Offers a more accurate and efficient way to evaluate ANN search, potentially speeding up development and deployment of AI systems reliant on it.