They’re likely using an HNSW index, which typically requires a lot of memory for large data sets.