Embedding visualizer

What it does

It projects vectors onto their first two principal components and draws them: each vector a point, labelled if you gave it a label. Hover a point for its label and coordinates. The same points are listed in a table below the plot, and Copy gives every one as tab-separated rows.

Try it: three made-up clusters of words, or an embeddings response as OpenAI's API returns it.

What it reads

The same inputs as the cosine similarity calculator: a JSON list of vectors or of {label, vector} objects, CSV or TSV rows with a label first, or an embeddings API response from OpenAI (including base64) or Gemini, pasted as it came back.

Up to 5,000 vectors, and up to 4,000,000 numbers in all — 5,000 vectors of 768 dimensions, or about 1,300 of 3,072. A large set is projected in a Web Worker, so the page keeps responding.

How it projects

PCA — principal component analysis:

1. Each dimension is centred on its mean.
2. The two directions along which the vectors vary most are found by block power iteration, with a Rayleigh–Ritz step each round. The covariance matrix is never built, so a thousand dimensions cost no more than they must.
3. Each vector is drawn at its coordinates along those two directions.

The axes say how much of the total variance each explains: a component's eigenvalue ÷ the total variance. The start is fixed, so the same input always gives the same picture, and each axis's sign is fixed by its largest coordinate. When two components explain almost the same variance, as in vectors with no structure, the iteration may not settle within its budget; the page then says the picture is approximate.

How to read it

Two dimensions keep what varies most and flatten the rest. Points drawn close together may still be far apart in the full space, especially when the two axes explain a small share of the variance. Check a pair that matters with the cosine similarity calculator, which uses every dimension.

PCA is linear. t-SNE and UMAP, which bend the space to keep neighbours together, are not offered.

What it does not do

It draws vectors you already have; it does not make embeddings. Salyro does not serve embeddings, and nothing you paste is sent to it or anywhere else.