#vector-databases

Embedding databases with hundreds of millions of vectors are costly to serve. CompresSAE uses a sparse autoencoder to cut the footprint 10×+ with only a small hit to retrieval quality.