[QST] How to choose parameters for CAGRA?

I want to select a good choice of parameters that would give me high recall (>95%) for two large datasets that I have. One of them has 120M rows and the other dataset has 50M rows. I have access to an NVIDIA A100 GPU with 80G VRAM. I want to compute 50 neighbors for every data point and so far the only change I have made is use itopk_size=128 when searching. Since datasets are too large I am not able to perform a brute force search and calculate recall. I was wondering if there are any guidelines on choosing parameters.

Start by reviewing CAGRA's search-parameter guidance, focusing on itopk_size and the requested 50-neighbor search. Use the 120M- and 50M-row datasets with the available A100 as the evaluation context. Done means documenting practical parameter-selection guidance for reaching the requested recall when brute-force recall measurement is unavailable.