Reading about embeddings tells you that “similar things end up close together in vector space.” This page lets you check that claim yourself.
Below is a small, hand-built set of words, each described by five simple properties (how big it is, how warm, how dangerous, and so on) instead of the hundreds or thousands of dimensions a real embedding model would learn on its own. The named properties make the numbers explainable; dimensions learned by a real model usually do not have such clear labels.
This is a teaching aid, not a real embedding model. It uses the same vector operations that production systems use, including cosine similarity and nearest-neighbor ranking.
Try this:
- Change the axes to see the same words form different clusters. The plot shows two dimensions, but similarity still uses all five.
- Click a word to make it the anchor. Larger, darker points are more similar, and the list shows the five nearest neighbors.
- Compare two words directly to connect their feature differences with the cosine similarity score.
- Try the vector arithmetic demo to see a relationship deliberately encoded in this toy space. Real embedding arithmetic is model-dependent and is not always this clean.