3,685 passages · 5 companies · same MiniLM embeddings as the PCA viewer, different lens.
Reading guide. PCA (the flagship viewer) preserves the directions of greatest overall variance; t-SNE and UMAP instead work to seat each passage next to its true nearest neighbors, at the price of distorting long-range distances. The knobs matter: low perplexity / n_neighbors favors fine local structure, high values favor the global picture. Structure that survives every setting — like this corpus's prose–tabular split — is a property of the data, not of the lens.