Current state-of-the-art vector approximate nearest neighbor search (ANNS) libraries mainly focus on how to do fast high-recall search in memory. However, extremely large-scale vector search scenarios present certain challenges. For example, hundreds of billions of vectors coupled with limited memory creates a capacity issue. There is also a scalability issue because increasing the number of serving machines increases query latency and computation costs. This occurs as a result of the search being done in each machine, and latency increases with the increased number of aggregating candidates. To address these challenges, we propose SPTAG++, a distributed ANNS system. In this talk, we’ll discuss SPTAG++, which is now integrated into production to support hundreds of billions-scale vector searches in production with millisecond response time and more than ten thousand queries per second.
Learn more about the 2021 Microsoft Research Summit: https://Aka.ms/researchsummit