Corey J. Nolet
Corey J. Nolet
This should also include moving kmeans over from RAFT, since it depends on the distances and 1nn APIs. Ideally, we would have a single kmeans API with an option for...
We need a scalar quantization API that is decoupled from the IVF APIs, but which can ultimately be composed into an IVF. Ideally, we would follow a UX similar to...
We need a separate product quantization API that is decoupled from IVF but can still be composed into IVF. Ideally this API would follow FAISS or Scikit-learn'a transformer estimators.
This is an important data processing technique that enables sparse vector search. We should support this in cuVS. Like all the other algorithms in cuVS, this API should be composed...
BL25 preprocessing, when used in conjunction with sparse vector retrieval, enables hybrid search. We should support this algorithm in cuVS. Like all the other algorithms in cuVS, we should be...
We fully migrated the pairwise distances C++ code from RAFT but we still need to expose C, Python, and Rust APIs for it.
The current problem with our pre-filtering search functions is that the predicate function is a template argument, which means each unique pre-filtering predicate function requires its own instantiation. This means...
We're removing the current static assertion that's being done because the actual struct for the ANN indexes doesn't align completely w/ the serialized data format and this is causing the...
For consistency, we should consolidate these. There's nearly no cost for having less headers in cuvs so we should make things as easy and straightforward for the user as possible.
The index base class is not just for ANN, and so we should namespace it appropriately. I would also like for the contents of ivf_list.hpp, Ann.hpp, and dataset.hpp to be...