while building the index the diskann paper explains about merging the edge sets of all vamana index graphs after they are built. why there is a need for merging the edge sets? Merging will not create a big graph of all 1B points in memory? What will be the output graph after merge? Like search, during indexing also we will find the closest centroid and add the new vector into that shard? During search, we find the closest medoid and execute the search in the vamana index related to that medoid.
while building the index the diskann paper explains about merging the edge sets of all vamana index graphs after they are built. why there is a need for merging the edge sets? Merging will not create a big graph of all 1B points in memory? What will be the output graph after merge? Like search, during indexing also we will find the closest centroid and add the new vector into that shard? During search, we find the closest medoid and execute the search in the vamana index related to that medoid.