Writing

A single dense vector quietly breaks down on rich items and compound user profiles. Tested across a full NBA season, late interaction (still rare in recsys) is the only retrieval method that holds up.

Why I'm excited about TopK's semantic_index: most retrieval systems buy quality by trading away freshness. Co-designing the model, inference engine, and database lets you keep both.

The LIMIT benchmark is held up as evidence that multi-vector models struggle. I tried to reproduce it — and found late interaction is embarrassingly good at exactly this task.

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.

Instagram's ranking changes are marketed as a win for small creators. Underneath, they read like a scalability play for early-stage collaborative filtering.