The simple explanation
An embedding model converts an input — text, an image, a snippet of audio — into a vector: a list of numbers, often hundreds or thousands long. Inputs that appear in similar contexts end up close together in that number-space. Search systems then compare vectors to find candidates with similar meaning even when the words differ: sofa and couch sit near each other; so do car and automobile. “Near” is defined by a chosen distance measure, and different systems measure it differently.
See this idea move.
The Embeddings Explorer experiment walks through this concept step by step — press run and watch it happen. Everything is simulated in your browser; no real AI runs.
A concrete example
The Embeddings Explorer draws a small map using hand-assigned food and transport scores. Move the query point and watch which items land closest — that’s similarity search in miniature. Real embeddings usually have far more dimensions and are learned from data, not assigned by hand, but the idea is identical: turn things into numbers, then compare by distance.
Those learned positions come from training: the model sees enormous amounts of text and gradually adjusts its vectors so that items appearing in similar contexts drift toward each other. Nobody programs “sofa is like couch” — the geometry emerges from the data. That’s also why embeddings inherit the biases of their training data: whatever associations were common in the text show up as closeness in the numbers.
Where people get misled
Near is a clue, not a verdict. Words that appear in similar contexts can be opposites — hot and cold live near each other in many embeddings. Nearness says nothing about truth: a false statement can sit right next to a true one. And embeddings don’t transfer between models — vectors from one model are meaningless numbers to another, and the distance measure changes which results count as “close.”
The honest limits
The model, the distance measure, and the task all affect which results are useful. Vectors can lose things the text had — negation, word order, precise quantities — so embedding search is a candidate-finder, not an answer-checker. Use it to find what might be relevant; verify from the source.