RAFT contains fundamental widely-used algorithms and primitives for machine learning and information retrieval. The algorithms are CUDA-accelerated and form building blocks for more easily writing high performance applications.
This PR updates the CAGRA test to normalize the dataset and query vectors in the CAGRA test when the metric is InnerProduct. If we don't normalize them, large L2 norm dataset vectors tend to be included in the search result across all queries. This means that only a part of the graph nodes may be traversed in the search process, leading to test incompleteness.
This PR updates the CAGRA test to normalize the dataset and query vectors in the CAGRA test when the metric is InnerProduct. If we don't normalize them, large L2 norm dataset vectors tend to be included in the search result across all queries. This means that only a part of the graph nodes may be traversed in the search process, leading to test incompleteness.