Star Wars Atlas: checking the map against the vectors

The input is fifteen authored character descriptions. They mention factions, roles, Force sensitivity, eras, and relationships. They are the dataset: a vector describes this wording, not an independently verified model of a character's entire history.

The published comparison tested nine embedding models. MiniLM emphasized the Skywalker family, while E5-large-v2 gave the strongest overall group separation in that comparison. Choosing E5 on these same descriptions makes this an exploratory example, not a held-out benchmark of embedding quality.

15 descriptions → passage prefix → E5-large-v2 → unit vectors
    → 15 × 15 cosine matrix → thresholded similarity graph
    → dimensionality reduction → labeled interactive map

E5 inputs use passage: and normalized embeddings. For unit vectors, the dot product equals cosine similarity. There are 15 × 14 / 2 = 105 distinct unordered pairs, excluding self-comparisons. That denominator is useful when interpreting graph density.

At a threshold of 0.87, 17 pairs survive, about 16.2% of the possible edges. Reusing MiniLM's 0.45 threshold would connect every E5 pair in this dataset. E5 scores lie in a compressed band around 0.80 to 0.94; a threshold is a model-specific choice, not a universal definition of relatedness.

The Jedi/Sith means do not support two cleanly separated camps. Palpatine is close to Mace Windu, whose description explicitly mentions confronting him. Five descriptions contain "Master", and those characters share a neighborhood. That is a concrete reason to inspect the input text when a cluster seems surprising.

Removing named relationships changed the result: the published ablation reports Sith-to-Sith similarity of 0.898 and Jedi-to-Sith similarity of 0.874. Vader-to-Palpatine rises to 0.910. These are results from edited inputs and must remain separate from the original matrix.

The illustrated UMAP configuration uses two components, five neighbors, minimum distance 0.3, cosine distance, and seed 42. Faction colors are supplied by the author. They are not labels discovered by the embedding model.

A two-dimensional arrangement cannot retain every relationship among 1,024-dimensional vectors. The article reports that two PCA axes preserve about 27% of variance; that PCA figure does not measure UMAP's retained variance. Use the map to select a pair, then inspect its original cosine value and descriptions. Screen distance alone cannot validate a claim about the embedding geometry.

The technical article includes all fifteen descriptions and code for encoding, graph construction, and projection. Preserve their exact wording and prefixes, model identifier, normalization, and projection settings. Check unit norms, matrix symmetry, and the unit diagonal before counting off-diagonal pairs.

Recompute the 105 pair scores and the 17-edge threshold result before interpreting clusters. Then repeat with relationship phrases removed, retaining both matrices. Changing the text and the model simultaneously prevents a useful attribution of the difference.

The recorded explorer demonstrates the interface. It is not a new replication of the model comparison. A stronger study would add unseen descriptions, repeated projection seeds, and sensitivity to wording. For this small dataset, the useful result is that relationship language and rank explain similarities better than the expected light-side/dark-side division.