Various data types can be mapped into high-dimensional embedding vectors through deep learning models. Being able to query the closest vectors to a given query vector enables semantic search, which has many applications in recommendation systems and machine learning. Because an exact solution is computationally infeasible, many approximate nearest-neighbor search (ANNS) systems have been developed, out of which the most promising approaches are graph-based. This project aims to improve previous ANNS systems and build a graph-based vector database that scales to billion-scale datasets. We will also explore variants of the classic ANNS problem, such as filtered search and out-of-distribution queries, and attempt to build a unified system that supports all of these cases efficiently.

Through this SuperUROP, I hope to apply my background in theoretical algorithms (e.g. from 6.5210 Advanced Algorithms) and C++ programming to a long-term research project with real-world applications. I hope to learn more about performance engineering and parallel computing as well as develop valuable skills required in academic research.