Embedding Vector Visualizer
Paste embedding vectors and see them plotted in 2D using PCA.
What is the Embedding Vector Visualizer?
Embedding vectors from language models typically have hundreds or thousands of dimensions — far too many to look at directly. This tool projects your vectors down to 2 dimensions using PCA (principal component analysis) and plots them as a scatter chart, so you can visually check whether similar items actually cluster together.
How the projection works — and its honest limits
This tool centers your vectors and finds the two directions (principal components) that capture the most variance in your specific data, using a matrix-free power-iteration method that works efficiently even for high-dimensional embeddings. It was verified against synthetic clustered data before release to confirm well-separated clusters stay visually separated after projection.
That said, PCA is a linear projection: it's excellent at revealing broad structure (well-separated clusters, dominant axes of variation) but it necessarily discards information, and two points that look close in the 2D plot are not guaranteed to be close in the original high-dimensional space (and vice versa) — especially for embeddings whose true structure is more complex than two dominant directions. For a more faithful (but heavier) non-linear alternative, dedicated tools like t-SNE or UMAP are worth exploring for serious analysis; this tool is meant for a quick, zero-setup sanity check, not a publication-grade visualization.
How to use it
- Paste your vectors, one per line, each with the same number of dimensions. An optional label before a colon (like
cat: 0.1, 0.2, ...) is used to annotate the plot. - Click Plot in 2D.
- Look for clustering: points that land near each other in the plot share similar directions in the original embedding space.
Everything runs locally in your browser using plain JavaScript — your vectors are never uploaded anywhere, which matters if they're derived from proprietary or sensitive text.