Embedding explorer
Add phrases, see where they land, and read the similarity between them.
Type phrases and watch them find their place in a live 2D projection.
What is real here
These are lexical vectors, not neural embeddings, and that difference is the whole lesson. Character n-grams, TF-IDF weighting, cosine similarity and PCA down to two dimensions — all of it computed in this page. Add a word and its synonym and watch them stay far apart: nothing here can see meaning, only spelling. Closing that gap is exactly what the dense embeddings inside this site's retriever do, and it is why the semantic search demo beats the keyword one. This API publishes no embedding endpoint, so instead of inventing one, the demo shows you the honest lower bound.
Each phrase is one point. Phrases built from similar characters sit close together.