UMAP projection of 1,336,165 documents from the EFTA production (DOJ Epstein Files Transparency Act release), House Oversight Committee, DOJ-OGR, FBI Vault, and USVI productions. Documents that share text content or are linked by the relation graph (email threads, near-duplicates, image-near-dup, containment, and verbatim-phrase edges) tend to land near each other.
Coordinates have no semantic meaning beyond "closer = more similar" — only local distances are reliable. Empty/blank docs (~121K with no extractable text) are excluded.
Use the filter bar (Type, Source, date range, has-X flags) or the text search box to subset the visible dots. Filters compose: text query intersects with all checked filters. Non-matching dots disappear; the rendering is otherwise unchanged.
Text search runs against the full FTS5 index (~200K result cap, usually plenty). All other filters run instantly client-side over the in-memory doc metadata.
Built on BGE-base-en-v1.5 embeddings of all document text, UMAP for 2D positions (n_neighbors=15, min_dist=0.05, pca init, fused with relation-graph edges).