Catalyst
Parquet-Based ETL Engine
The 3D Semantic Ziggurat is not an artistic rendering. It is a deterministic, relational 3D topography computed directly from Lakehouse vectors, cluster trees, and sibling-relative chunk counts. Raw document payloads transform into stepped physical geometry without manual intervention.
Each document chunk is embedded using Model2Vec, distilling transformer embeddings into dense 256-dimensional static vector arrays. In-process execution completes at over 50,000 sentences per second on Apple Silicon.
Uniform Manifold Approximation and Projection (UMAP) constructs fuzzy simplicial sets across the 256D semantic space, projecting embeddings onto a 2D continuous metric plane while strictly preserving local neighborhood and global topological distance.
HDBSCAN identifies density-based clusters without assuming spherical shapes. It generates a multi-tier cluster hierarchy spanning Depths 1 through 4, partitioning the corpus into macro-continents, territories, and topical neighborhoods.
Planar Voronoi tessellation assigns discrete geometric boundaries to cluster centroids. Smooth cubic Bézier curves trace continental coastlines and territorial divides, establishing zero-overlap spatial footprints across the bedrock plane.
Child clusters stack directly upon their parent terrace decks. The vertical elevation of each tier scales with sibling-relative evidence chunk volume: high-density subject domains rise into towering ziggurat peaks, while peripheral topics form broad lowlands.
The Sublate 20-mineral palette anchors to Lapis Blue (210°) at the westernmost continental centroid and sweeps counter-clockwise around the center of mass. Calibrated PBR shading depth-attenuates albedo from bedrock (0.22) to summit pavilions (0.90).