Music embedding map
Each dot is one chunk from the Chroma collection, laid out in 2D with t-SNE from its 1,536-number embedding. Chunks the model sees as similar sit close together. Distances between clusters are distorted, so read the map for who sits with whom, not for how far apart groups are.
The query star tends to sit at the edge of the cloud, away from what it retrieved. A short question is less similar to any chunk (the best cosine distance is often around 0.5–0.6) than chunks of the same artist are to each other, so t-SNE has no close neighbours to place it among. The lines show what was retrieved; the table below gives the true distances.
Retrieved chunks
| # | Artist | Chunk | Cosine distance |
|---|
Nearest neighbours in the full 1,536 dimensions
Checked on the original vectors, not the 2D map, so t-SNE can't distort it. For each chunk: is the single most similar other chunk from the same artist, and of the same kind?