Embeddings & Vector Search, Explained From Zero

Almost every impressive AI application — the assistant that answers from your company's documents, the tool that remembers you, the search that just gets what you meant — rests on one idea that sounds abstract but is genuinely simple once you see it. This guide starts from nothing and builds up that idea: how a computer can understand what text means well enough to find related things, even when they share no words at all.

1 · Why keyword search misses the point

For decades, computers searched by matching words. You type “car,” it finds documents containing “car.” Simple — and frustratingly literal. Search “car” and a perfect article that only says “automobile” or “vehicle” is invisible, because it never used your exact word. The computer wasn't matching meaning; it was matching letters. The deeper problem is that meaning and spelling are different things: “how do I fix a flat tire” and “repairing a punctured wheel” mean nearly the same thing while sharing almost no words. Keyword search can't bridge that gap, and for a long time neither could computers. Embeddings are how that finally changed.

2 · A map where meaning has a location

Here's the mental picture to hold onto. Imagine a giant map, and every possible idea has a spot on it. Ideas about dogs cluster in one region; ideas about cars sit in another, far away. Crucially, closeness on this map means closeness in meaning: “puppy” lands right next to “dog,” while “bulldozer” sits far from both. On a map like this, finding related things stops being about shared letters and becomes about distance — what's near this idea? “Car,” “automobile,” and “vehicle” would all be neighbours, because they mean nearly the same thing, regardless of how they're spelled. This map of meaning is the whole concept. Everything else is just how we build it and use it.

3 · An embedding is coordinates for meaning

So how does a computer place text on that map? On any map, a location is a set of coordinates — a few numbers that pin down a point. An embedding is exactly that: an AI model reads a piece of text and outputs a list of numbers that are its coordinates on the map of meaning. Feed in “the dog barked” and you get back its position; feed in “the puppy yelped” and you get a position very close by, because the meanings are close. The numbers themselves aren't meaningful to read — you'd never look at them and see “dog.” What matters is the relationship between points: texts that mean similar things get similar numbers, so they land near each other. Turning text into these coordinates is called embedding the text, and the model that does it is an embedding model.

4 · Far more than two directions

A real map has two directions — north/south and east/west — so two numbers fix any point. Meaning is far richer than that, so an embedding uses many more numbers, often hundreds or thousands. Each is like an extra direction the map can stretch in, letting it capture countless shades of meaning at once — one axis might roughly track “animal-ness,” another “size,” another “formality,” and so on, though in practice these directions aren't cleanly labelled. You can't picture a thousand-dimensional space, and you don't need to: the intuition is identical to a 2-D map, just with far more room, so subtle differences in meaning each get their own bit of space. More directions simply means a more faithful map.

5 · Vector search: find the nearest neighbours

Now the payoff. Once every text is a point on the map, searching becomes beautifully simple: embed your question into a point too, then find the points closest to it. Those nearest neighbours are the texts most similar in meaning to what you asked — and that's your search result. This is vector search (a vector is just the technical word for that list of coordinates). Ask “how do I fix a flat tire” and the system lands your question on the map, looks around it, and finds the article about “repairing a punctured wheel” sitting right nearby — even though the words are completely different. It found the answer by meaning, which is exactly what keyword search never could. Search stopped being letter-matching and became neighbour-finding on a map of meaning.

6 · Where the map lives: the vector database

To make this practical at scale, you need somewhere to keep all those points and search them fast. That's a vector database: a store built specifically to hold millions of embeddings and, given a new query point, quickly return its nearest neighbours. You embed all your content once, up front, and load the points into the database; from then on, every search is a fast “what's nearest to this?” lookup. Alongside each point, the database keeps the original text and useful labels — which document it came from, when, who's allowed to see it — so results can be shown, cited, and access-controlled. Building this store is a one-time cost; searching it is cheap and instant. The vector database is simply the map of meaning, made real and searchable.

7 · Why this quietly powers everything

Embeddings feel abstract until you notice how much they hold up. When an AI answers questions from your company's documents, the first step is embedding those documents and vector-searching them to find the relevant passages — that's retrieval, the next topic. When an assistant “remembers” you across conversations, it often stores and retrieves memories the same way. Recommendation, deduplication, clustering, semantic search — all lean on the same trick of turning meaning into coordinates and measuring distance. So this isn't a niche technique; it's the foundation layer. Understand embeddings and vector search, and a whole tier of modern AI apps stops looking like magic and starts looking like a sensible, understandable machine: turn things into points, then find what's near.

8 · A simple test you can run this week

9 · Glossary — every term, spelled out

Embedding

A list of numbers that represents a piece of text's meaning as coordinates on a map, so similar meanings get similar numbers.

Embedding model

The AI model that reads text and produces its embedding.

Vector

The technical name for that list of coordinates — a point in the space of meaning.

Vector search

Finding the texts whose points are closest to a query's point — searching by meaning, not by matching words.

Nearest neighbours

The points closest to a given point; in search, the most similar-in-meaning results.

Vector database

A store built to hold many embeddings and quickly return the nearest neighbours to a query.