Skip to content
latentSource

Project showcase

Star Wars embedding atlas

An interactive explorer for fifteen character descriptions, with similarity scores, projection views, and a way to test what changes when you remove their named relationships.

Sentence embeddings · E5-large-v2 · PCA · UMAP

Read technical article
Map characters through their descriptions

Recorded demonstration. Unmute the player to listen.

What it does

I built the atlas to check the clusters I had described in the article. Selecting a character shows its source text and nearest neighbors. Map, graph, and matrix views let me inspect the same scores from different angles.

Switching to the relationship-free descriptions reveals how much the family tree and named connections influence the results. The app shows the text changes alongside the recomputed rankings. Its embeddings are prepared offline; exploring the atlas makes no model calls.

The recorded walkthrough uses the working app, with English AI narration. It follows Luke's neighbors, the bounty-hunter pair, graph thresholds, and Vader's changing similarity to Palpatine.

How it is built

Read the technical post for the embedding pipeline, model comparison, and the cluster claims I corrected after checking the scores.

Mapping the Star Wars Universe