The match that wins, dry with clear skies, shares not one word with your query. Keyword search ranks a rock song called purple rain first because the letters line up.

Every document and every query word is placed along the same handful of concept axes. Two lines about money point the same way even with no shared words, so the cosine between them is near one. A weather query and a cooking note point in different directions, so their cosine is near zero. Direction is the meaning.

Embeddings put meaning into geometry: similar ideas become nearby directions, unrelated ideas become perpendicular ones. Cosine similarity reads only that direction, ignoring how long a vector is, so it asks do these mean the same thing rather than do they share the same letters. That is why a paraphrase or a synonym still lands on the right document, and why a coincidental word match gets pushed down where it belongs.

Keyword search has the opposite failure: it can only rank what it can spell. When your words and the answer's words diverge, it returns the wrong line or nothing at all. Meaning fills exactly that gap.

These vectors are hand placed along seven named topics so you can read the geometry directly. Real systems learn the embedding from text with a neural model, and the space has hundreds or thousands of dimensions with no human-readable labels. The cosine ranking math is exactly the same; only the source and size of the vectors change. At scale the query embedding is compared against millions of document embeddings in a vector database, and feeding the top matches to a language model is what retrieval augmented generation, RAG, actually does.