Open cjnolet opened 5 months ago
FAISS has a CPU API for merging IVF indexes, wherein they simply add the IVF lists from one index to the other and reset the indices of the merged lists so that there are no repetitions. This is somewhat easy to implement for our IVF indices too. Will be more complicated to merge optimized CAGRA graphs though.
A common operation that's needed from some vector database architectures is the ability to merge two or more indexes into a single larger index. This is something that we should abstract behind a cuVS API so that we can continue to improve its performance over time.
(cc @benfred @tfeher @akiranaruse as I know we've discussed this in the past)