Hadrons inherit the direction of their parent parton; anti-kT reconstructs the parton from the spray. Documents inherit the direction of their parent idea; k-NN reconstructs the idea from the corpus. Two detectors, one phenomenon: clustering by angle. Click the dictionary below to compare instruments — or press ▸ cluster on both sides and watch the same algorithm work both universes.
Smash two nuclei; a quark flies out. You never see it — it instantly dresses itself in a spray of particles flying the same way. That spray is a jet: the reconstructed answer to the query “what just happened in there?”
An embedding model turns text into a vector; texts about the same thing point the same way. A query is a probe vector, retrieval collects everything within a small angle. It is jet finding in meaning-space, with worse units and no peer review.
standing byPick an entry. The detectors will translate.
In this demo, both panels really do run the same code: one sequential-recombination routine, one exact k-NN. In production they don't — jets use FastJet, retrieval uses HNSW (approximate: it trades recall for speed, and yes, that's measured). Same mathematical family — agglomerative clustering and neighbor queries in an angular metric — different species.
Clustering ≠ retrieval. Jet finding partitions the whole event; search answers one query. The kinship is the metric and the computational-geometry toolbox (both fields escaped O(N²) with Voronoi diagrams and friends), not a shared codebase.
2D here, ~3072-D in reality. High dimensions concentrate angles — everything becomes nearly orthogonal to everything, which is precisely why approximate indexes exist.
“Direction = meaning” is learned, not conserved. Momentum conservation is a law; embedding geometry is a habit the model picked up in training. There is no Noether theorem for semantics. Yet.