Various data types can be mapped into high-dimensional embedding vectors through deep learning models. The ability to query the closest vectors to a given query vector enables semantic search, which has numerous applications in recommendation systems and machine learning. Since an exact solution is computationally infeasible, many approximate nearest-neighbor search (ANNS) systems have been developed. However, real-world applications may impose structured label constraints on the desired search results, a variant of ANNS known as filtered search. Despite the smaller search space, it is challenging to integrate filters with vector search efficiently, and existing methods waste a lot of computation processing irrelevant vectors. This project aims to improve previous filtered search systems and build a vector database that scales to billion-scale datasets.

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.