Description
Right now, for the brute-force recall, we are using a vector scorer source. Which pulls vectors on heap, uses Lucene similarity function, and captures the nearest K.
This bypasses all our optimizations for bulk off-heap scoring.
I think we can improve the brute-force scoring costs significantly through: #147820 once its merged.
That query should allow for fast off-heap bulk scoring, and shouldn't be that hard to integrate in the KnnIndexTester.
Description
Right now, for the brute-force recall, we are using a vector scorer source. Which pulls vectors on heap, uses Lucene similarity function, and captures the nearest K.
This bypasses all our optimizations for bulk off-heap scoring.
I think we can improve the brute-force scoring costs significantly through: #147820 once its merged.
That query should allow for fast off-heap bulk scoring, and shouldn't be that hard to integrate in the KnnIndexTester.