Embedding Projector
Paste labelled vectors and project them to a 2D scatter with principal component analysis to see how they cluster, then pick any point for its nearest neighbours by cosine similarity. Everything runs in your browser.
New to this? Read the embedding projector guide →
PCA reduces your vectors to two dimensions for display — some information is always lost, so read clusters, not exact coordinates. Everything runs locally in your browser.
How to use the embedding projector
- Paste your vectors, one per line, each as a label followed by its numbers — cat: 0.9, 0.1, 0.0. All vectors must have the same length.
- Press Project to 2D. The tool runs principal component analysis (PCA) and plots each vector as a point; similar vectors land near each other.
- Pick a point in the sidebar to list its nearest neighbours by cosine similarity — the same measure most vector databases use for retrieval.
What PCA is doing
Embeddings often have hundreds or thousands of dimensions, which you cannot see. PCA finds the two directions along which your points vary the most and projects everything onto them, giving a flat map that preserves as much of the spread as possible. Points that sit together on the map are usually close in the full space too — but because you have thrown away every other dimension, treat the picture as a sketch of the structure, not a precise measurement.
PCA vs cosine similarity
The scatter shows the overall shape; the neighbour list gives the exact relationships. Cosine similarity, computed on the original full vectors, is what actually drives retrieval in a RAG system, so it is the number to trust when two points look close on the map. For more on that measure, use the cosine similarity calculator and read embeddings and cosine similarity explained.
Frequently asked questions
What does the embedding projector do?
It reduces high-dimensional vectors — word or sentence embeddings, feature vectors — to a 2D scatter using principal component analysis (PCA), so you can see which ones cluster together. Selecting a point lists its nearest neighbours by cosine similarity on the original full vectors.
How should I format the vectors?
One per line as a label then its numbers, e.g. "cat: 0.9, 0.1, 0.0". Commas or spaces both work, and every vector must be the same length; lines that do not match are skipped with a note.
Are the 2D positions exact?
No. PCA keeps the two directions of greatest variation and discards the rest, so distances on the map are approximate — read it for clusters and outliers, not precise measurement. When closeness matters, trust the cosine score, which is computed on the full vectors.
Is my data uploaded?
No. The PCA and similarity maths run entirely in your browser, so you can visualise embeddings of proprietary text without anything leaving the page.