React Embedding Space Map
A 2-D projection of an embedding space, so a human can see what a model considers similar. Clusters show the structure it found; selecting a point draws its nearest neighbours — which is how you check a retrieval or classification system by looking, not by reading cosine scores.
- Charts & analysis
- Forms & inputs
- Semiconductor
- AI & agents
Select a point to see its nearest neighbours.
Hover or click a point to draw its nearest neighbours. The clusters are the structure the embedding found; the neighbour links are how you check it makes sense.
Installation
Props
| Prop | Type | Description |
|---|---|---|
| points | EmbeddingPoint[] | { id, x, y, cluster, label } — x/y in 0–1 |
| clusters | EmbeddingCluster[] | { id, label, tone? } |
| k | number | Neighbours to draw on select (default 4) |
See the structure, check the neighbours
Embeddings are high-dimensional and unreadable until you project them; once projected, a glance tells you whether the model's notion of similarity matches yours. The clusters render the grouping the embedding found, and selecting a point draws links to its k nearest neighbours with their distances — the concrete answer to "why did retrieval return these?" A neighbour from the wrong cluster is a visible, clickable bug report.
Projection is yours; exploration is here
The component does not compute the projection — that is UMAP, t-SNE, or PCA run in your pipeline, and it belongs there. What it gives you is the interaction: hover to select, links to neighbours, distances, a cluster legend, and the dimming that keeps a busy map legible. Feed it normalised 0–1 coordinates and cluster ids, and it turns a static scatter into something you can interrogate. Pair it with Semantic Search Results to go from a scatter of near points to the ranked list a query actually returns.
New components every week
Get the week's new Kelvin UI components and templates in one short email. No spam, unsubscribe anytime.