Computers are very good at matching letters and very bad at matching meaning. “Two-wheeler parking” and “bike parking” are the same thing to any person and completely different strings to a search box. Embeddings are how that gap gets closed.
Here is the idea in one picture. Imagine a giant map, and every word or sentence in the world has a place on it. Things that mean similar things are placed near each other: “dog” sits close to “puppy”, a little further from “cat”, and a very long way from “mortgage”. Once everything has a position, “what is this similar to” becomes a question about distance, and distance is something a computer can measure perfectly.
An embedding is just that position, written down as a list of numbers.
In short
- What
- A list of numbers that stands for a piece of text or an image, positioned so that things meaning similar things end up near each other.
- Why
- It turns “does this mean roughly the same thing” — which computers cannot do — into “how far apart are these two points”, which they do instantly.
- When
- Searching by meaning rather than by exact words, grouping similar items, spotting near-duplicates, or routing something to the right category.
- How
- Split the text into pieces that each make sense alone, embed them, keep the original text beside the numbers, and test against real queries rather than tidy ones.
Why numbers, and why so many
A colour is a familiar version of the same trick. Any colour can be written as three numbers, and once it is, you can measure how close two colours are without knowing what either is called. Nobody had to list every shade and its neighbours; the numbers carry the relationship.
Meaning needs far more than three numbers, because meaning has far more ways to vary — formal or casual, about money or about places, a question or a statement, and hundreds of other dimensions nobody has to name. So an embedding is typically a list of several hundred or more numbers, each one a fraction of a direction on that map.
No single number means anything on its own. It is the whole arrangement that puts “bike parking” near “two-wheeler parking”.
What this makes possible
Search that understands the question. Someone typing “place to keep my scooter” finds a listing that says “covered two-wheeler parking”, even though not one significant word matches. Keyword search cannot do this; it has no way to know the two phrases are related.
Grouping without being told the groups. Feed in a few thousand support messages and the ones about the same underlying problem cluster together, whether or not anyone anticipated that problem.
Catching near-duplicates. Two listings for the same home, written by different people in different words, sit close together on the map even though no field matches exactly.
None of this requires the system to understand anything. It only requires a good sense of what is near what.
When you reach for it
When the words people use will not match the words in your data. That is most of the time, once real people are typing.
When you have more material than anyone could organise by hand, and you need it grouped, deduplicated or routed.
And when you do not. If the thing you are matching is a code, an identifier, an exact name or a number, ordinary matching is faster, cheaper and exactly right. Embeddings are for meaning, and a postcode does not have any.
How to use them well
Split text into pieces that make sense on their own. A paragraph usually works; half a sentence does not. A piece that needs the one before it to be understood will be no use when it comes back on its own later.
Keep the original text beside the numbers. The numbers are for finding; the text is what you actually show or hand to a model. A system that stores only the embedding has thrown away the useful half.
Use the same model at both ends. Embeddings made by two different models live on two different maps, and distances between them mean nothing at all. Change the model and everything has to be made again — worth knowing before you have millions of them.
Test with real queries, including the badly typed ones. Embeddings look flawless on the examples you thought of and reveal their gaps on the ones you did not.
What they do not do
Similar is not the same as correct. The nearest passage to a question is often the most useful one, and sometimes it is merely the one that sounds the most like it. A confidently-worded wrong paragraph sits very near a question about the same subject.
They also have no sense of time or truth. An embedding of an out-of-date policy sits exactly as close to the question as the current one does. Keeping meaning and freshness straight is a job for whatever stores and filters them — usually a vector database — not for the numbers themselves.
