VectorChord has made significant advancements in indexing 100 million 768-dimensional vectors in just 20 minutes using PostgreSQL. This process previously required about 200 GB of memory and 40 hours with pgvector, making it impractical for large-scale applications. The new method reduces memory usage by seven times and builds indexes significantly faster, allowing deployment on cheaper machines without GPUs. The article details several optimizations across three phases: initialization, insertion, and compaction. In the initialization phase, hierarchical K-means clustering was improved to speed up the process from 30 minutes on a GPU to just 8 minutes on a CPU. This was achieved by dividing the data into smaller subsets for clustering, which allowed for a more manageable computational load. The insertion phase saw a dramatic decrease in time from 420 minutes to just 9 minutes, while the compaction phase was reduced from 8 minutes to 1 minute. Memory usage was also tackled by focusing on dimensionality reduction. The article references Christos’s findings that clustering can still be effective on lower-dimensional projections, which helps mitigate the high memory requirements of the original K-means algorithm. Overall, these optimizations enable VectorChord to manage substantial vector datasets efficiently while maintaining a balance between speed and accuracy.