What Are Vector Embeddings? Explained Simply

Keyword search has a blind spot: it matches letters, not meaning. Search your notes for "server down" and you'll miss the one that says "502 after the deploy," even though it's exactly what you wanted. Vector embeddings are how modern systems fix that — they let a computer find things by what they mean, not what they literally say. And the core idea is surprisingly visual.

The problem: computers see strings, not meaning

To a plain program, "car" and "automobile" are as unrelated as "car" and "banana" — just different sequences of characters. Anything that searches, groups, or recommends by meaning needs some way to represent that meaning as data a computer can compare. That's the job embeddings do.

What an embedding actually is

An embedding model reads a piece of text and outputs a vector — a list of numbers, often a few hundred to a couple thousand of them:

"502 after a deploy"  ──▶  [0.021, -0.114, 0.077, ... ]   (e.g. 1024 numbers)

Think of each vector as coordinates in space. The model is trained so that texts with similar meaning get similar coordinates — they land near each other — while unrelated texts land far apart. "The site is down" and "502 after deploy" end up as near-neighbours; "chocolate cake recipe" sits on the other side of the room.

Similarity is just distance

Once text is coordinates, "how related are these two things?" becomes "how close are these two points?" — a simple calculation (usually cosine similarity, the angle between the vectors). To search by meaning you embed your query, then find the stored vectors closest to it. No keyword needs to match.

An embedding doesn't store what the text says — it stores where the meaning sits. Search stops being "find these words" and becomes "find what's nearby."

Why it matters

This one primitive powers a lot of what feels like AI magic. Semantic search finds relevant docs by meaning. RAG ("chat with your documents") uses embeddings to pull the right passages before an LLM answers. Recommendations surface similar items. Clustering and deduplication group things that mean the same. Each is the same move: turn things into vectors, then reason about closeness.

The mental model to keep

Picture a giant map where every piece of text is a pin, and the model has arranged the pins so related meanings are physically close. Embedding is dropping a pin; searching is standing on your query's pin and looking at the nearest neighbours. You don't need to understand the geometry to use it — you need to embed your content once, store the vectors, and compare distances at query time. That's the whole engine under semantic search and RAG.

Frequently Asked Questions

What is a vector embedding?

A list of numbers that represents a piece of text's meaning, arranged so that texts with similar meaning get vectors that land near each other.

How are embeddings used in search?

Embed your content once and store the vectors, then embed the query and return the stored vectors closest to it — searching by meaning instead of keywords.

What is cosine similarity?

A measure of how closely two vectors point in the same direction; it's the usual way to score how related two embeddings are.