Martin Spišák
· 3 min read

The future is sparse: compressing embeddings with CompresSAE

sparse-embeddings recsys vector-databases

Vector databases are getting huge — and sparse embeddings are here to help.

White light passing through a glass triangular prism and splitting into a rainbow spectrum.
A sparse autoencoder is a prism: send a dense embedding through it, and its hidden structure splits out into a handful of meaningful dimensions.

Embedding databases with hundreds of millions of vectors are nothing new in industrial recommender systems, but serving them at that volume remains costly and strenuous. To ease that strain, our team at Recombee developed CompresSAE: a lightweight, highly scalable embedding compression technique based on a novel sparse autoencoder (SAE).

The idea in one paragraph

With CompresSAE, reducing your embedding footprint by more than 10× — with only a small trade-off in downstream retrieval quality — can be as simple as:

  1. Train a lightweight SAE on embeddings already sitting in your vector database.
  2. Use the trained SAE encoder to transform your dense embeddings into a compressed sparse format.
  3. Serve them with a vector database that supports sparse vectors, and retrieve with standard cosine similarity.

No exotic serving path, no new similarity function — the compressed vectors drop into the infrastructure you already run.

Why it matters

Our results show a favorable compression–quality trade-off in real production deployment. That’s the part I find most interesting: it challenges the default assumption that large-scale vector search has to mean dense embeddings. For a lot of systems, the dense vector is carrying far more bits than the retrieval task actually needs.

Left: a table showing CompresSAE delivering a higher CTR lift than SBERT and Matryoshka at a fraction of the storage. Right: a Recall@100 vs compression-factor plot where CompresSAE holds up better than equally-sized Matryoshka embeddings.
CompresSAE beats equally-sized Matryoshka embeddings on the compression–quality curve, and led to a CTR lift in an A/B test — measured with anytime-valid confidence intervals.

The same idea travels beyond retrieval. We later used the sparse-autoencoder trick to surface segment-level insights for real-time editorial support across a news media group — Segment-Aware Analytics for Real-Time Editorial Support in Media Groups, which builds directly on CompresSAE.

We presented CompresSAE as a Spotlight talk at RecSys 2025.

Big shout-out to the team at Recombee who made CompresSAE happen: Petr Kasalický, Vojtěch Vančura, Daniel Bohuněk, Rodrigo Alves, and Pavel Kordík.