The Semantic Tile

Concept · geography_cartography · hypothesis · visibility: publishable

A vector tile is pre-baked geometry plus attributes for one styling at one zoom — optimized for fast
display (binary-packed, generalized per zoom, CDN-cached) but dumb: its meaning is frozen at
tile-generation time, it cannot infer or recombine, and to change what it shows you re-tile. A
semantic tile inverts the priority: the tile’s contents are defined by a query over a triple store,
so the map is a query result, not a product. ‘Administrative boundaries’ vs ‘vernacular boundaries’
vs ‘figure-ground’ become different maps from the same substrate — ‘visualize space, not make a map’
made literal: the tile is a question. Figure-ground is the purest case, because the figure/ground
decision is itself a semantic class query (a region typed water becomes black, its complement white) —
the most semantic cartographic style there is, Stamen Toner as one SELECT.

The honest architecture, no hype: RDF is verbose and slow to render raw; vector tiles win the
painting race by design and always will. So the win is not ‘semantics renders faster’ — it is that
semantics is queryable and inferential where a tile is inert. The real lightness is not bytes
(triples are heavy); it is inference: a triple store derives transitive containment (A in B, B in C,
therefore A in C) without storing it, so you need not bake every derivable relation. A vector tile can
infer nothing. The buildable form is therefore not semantics-instead-of-tiles but semantics as
source-of-truth with a derived, cached render layer — materialised semantic tiles: tile-speed display,
query-defined contents.

Connections

source: live code session (2026-06-15) · extraction: auto (schema v3.1)