An embedding is a list of numbers that represents the meaning of a piece of text. A model computes it, and texts with similar meaning end up close together even when they use different words: "car" and "automobile" land side by side. That is why they power search by meaning and are the basis of RAG.
A question from the app. Pick an answer.
You search a store's help page for "broken phone" and get "my screen cracked". How?
They do not share a single word. It matched them by what they mean, not by how they are spelled.
An embedding model does not write: it turns each text into numbers, and texts with similar meaning end up close. The search returns the closest ones.
It is not luck: it measures how close each text's meaning is to your search and ranks by that.
You give a text to an embedding model, which does not write answers: it returns hundreds or thousands of numbers. None of them means anything on its own; what matters is where that list lands compared to others. Think of it as a map with a huge number of directions, where each text is a point.
To tell whether two texts are alike, you measure how close their points are. Close means similar meaning, even if they share not a single word. Far means they are about something else. That measurement is the whole trick.
Search by meaning: a store's help section finds the answer even when you phrase the question your own way. They are also used to group similar reviews, spot repeated questions or recommend related articles.
And they are the first step of RAG: your documents are split into chunks, each chunk becomes an embedding and, when a question comes in, the closest chunks are found and handed to the model.
In English, the same thing changes name depending on where you are: sneakers or trainers, apartment or flat, truck or lorry. A search that matches exact words gets lost there. One built on embeddings usually finds it anyway, and with a multilingual model a question in English can find a document in Spanish.
"The embedding understands the text." It captures what the text is about, not whether it is true, and it can place opposite sentences like "I like it" and "I do not like it" close together, because they are about the same thing. "Embeddings from one model work with another." No: if you switch models, you recompute everything. "It is the same as a token." A token is a piece of text; an embedding is the numbers that represent its meaning.