Live demo

Watch the lattice win - in your own browser

These are simplified JavaScript demonstrations of the real engine's structures, running on synthetic data on your CPU right now. The shipped engine is native code and much faster.

Synthetic data only. Your browser, your CPU. No external scripts, fonts, images, or network calls.

Demo A

Find a name among 1,000,000 files

Generate synthetic realistic file paths, build a simplified version of the LatticeSearch index, then race it against a full linear scan on the same query.

LatticeSearch engine vs linear scan

Builds a search-ready name index, then verifies candidates before returning hits.

Waiting to generate synthetic paths.

Linear hits

    LatticeSearch hits

      Demo B

      Find meaning among 100,000 vectors

      Build synthetic clustered 64-dimensional unit vectors in Float32Array, then race exact flat scan against a simplified lattice descent.

      Lattice engine vs flat scan

      Multiway spherical clustering, then query-time descent that scores only reached groups.

      Waiting to generate synthetic vectors.

      Demo C

      The Terza face-off

      LatticeSearch Terza is the upcoming business tier: a vector engine that narrows embedding space through layered descent. Terza — built to search less and still find the answer.

      Exact scan vs conventional index vs Terza engine

      Same synthetic clustered vector corpus as Demo B, three engines, one random query.

      Waiting to build the shared vector corpus and comparison engines.

      Simplified in-browser demonstration on synthetic data. The shipped Terza engine is native code with accuracy tuning that reaches ~95%+ agreement with exact search.

      Measured on metal

      The real engine, measured

      The browser demos above are simplified demonstrations on synthetic data. These are the measured numbers for the native engine on real data.

      WorkloadLatticeSearchBaselineResult
      Keyword lookup over 1,849,709 real files0.009-0.24 ms coreWindows Search 30-44 ms100-4,000× faster
      Semantic candidate lookup, Terza engine on real CLIP/e5 corporaTerza engine 1.1-2.3 ms at ~92% recallFlat scan 3.2-8.5 ms exact; conventional index at matched recallTerza beat convention ~25-30% at matched recall and built ~40% faster.
      Meaning query: "screenshots of a login page"46 resultsWindows: 0 resultsMeaning search finds what filename search cannot.
      Usual searches after resuming a project~7 msWarm Working SetFast enough to stay in flow.

      Measured 2026-07-21 on one Windows 11 machine, medians; methodology in the comparison report; your hardware will vary.

      Proof you can run

      You can't deny a benchmark that runs on your own machine.