The primitive that retrieval is built on, made visible. Seventy-seven EA and AI concepts projected from 1024-dimensional embedding space down to a 2D map via UMAP. Type a query — your text gets embedded live, its five nearest neighbors highlight, and the query lands near them. Hover any point to see its source and similarity.
Place a query to see telemetry.
The map you see is a 2D UMAP projection of the 77 corpus embeddings, computed offline at corpus-build time. UMAP doesn't trivially extend to new points — to project your query through the same model, you'd need to retain the trained UMAP transformer, which is finicky to ship to the browser.
What this demo does instead: embed your query in the original 1024-dim space (server-side, one NVIDIA call), find the 5 nearest neighbors by cosine similarity (client-side), then plot your query at the centroid of those neighbors' 2D coordinates with a small jitter so it doesn't stack on top of an existing point.
This is a placement approximation, not a true re-projection. The neighbor-distance scores are accurate (cosine on the original 1024-dim vectors); the visual position relative to non-neighbors is "near where its closest matches live, plus jitter." Honest framing matters more than pretty visuals — a real-world version of this demo with a stable transformer would just re-run UMAP.transform() on the new point.
RAG Explorer shows the application: a question goes in, a grounded answer comes out. Embeddings Atlas shows the primitive: how meaning is encoded as geometry. Together they explain why retrieval works — semantically similar concepts cluster — and why it sometimes doesn't, when the question lives in a sparse region of the map.