Martin Spišák · Applied Research @ TopK
Hi, I'm Martin.
I make retrieval scale.
Search research at TopK, previously RecSys at Recombee. I build the algorithms that make retrieval scale — cheaper, faster, fresher. It's the layer underneath nearly all search and recommender systems, so the wins compound everywhere.
Topics
Writing
all posts →- We don't need to compromise on freshness in retrieval
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.
- Late interaction has no LIMIT
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.
- The future is sparse: compressing embeddings with CompresSAE
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.
- Your turn, TikTok: the hidden reason Instagram's new algorithm could win short video
Instagram's ranking changes are marketed as a win for small creators. Underneath, they read like a scalability play for early-stage collaborative filtering.
Elsewhere
- → High-Quality Search, Out of the Box topk.io TopK's semantic_index: production multi-vector retrieval from a single schema annotation.
- → SMVE: Multi-Vector Retrieval That Just Works topk.io Turning late-interaction (ColBERT-style) embeddings into compact sparse vectors so multi-vector retrieval actually scales.
- → Iso-ModernColBERT huggingface An isotropically-corrected ColBERT that recovers sparse first-stage retrieval quality under SMVE.
- → RecSys 2025 Paper Explorer recsys25.recombee.com Personalized search and recommendations over the entire RecSys 2025 program — built on Recombee.
- → The Future is Sparse — talk youtube My RecSys 2025 talk on compressing dense embeddings into sparse ones for scalable retrieval.
- → SANSA — talk youtube My RecSys 2023 talk on SANSA: scaling EASE to millions of items with sparsity.
- → SANSA — code github Sparse EASE for millions of items — still running retrieval over billions of interactions daily.