Finding the closest items in huge datasets is slow because comparing every item takes too long. Approximate nearest neighbor search speeds this up by using clever indexes that find good matches quickly, but those indexes can still be slow and use a lot of memory.
The paper presents Ginger, a new C++ library that builds better indexes for this search. It uses a technique called Nested TurboQuant to store vectors using far less memory while keeping high accuracy, and it compresses the graph structure that guides the search without slowing it down.
At a target accuracy of 95%, Ginger answers queries about 1.8 times faster than the previous fastest library across many different datasets. This means applications like recommendation systems and image search can return results faster while using less computer memory.