Computers only work with numbers, so before an AI model can do anything with words, images, or other content, that content has to be converted into numbers. The specific technique used for this in modern AI is called an embedding.
What an embedding actually is
An embedding is a long list of numbers (a "vector") that represents the meaning of a piece of content, positioned in a huge mathematical space. The key property that makes embeddings useful: things with similar meaning end up with similar numbers, and therefore end up positioned close together in that space.
For example, the embeddings for "puppy" and "dog" would land close together, while the embedding for "spreadsheet" would land far away from both — even though none of those words share any letters in common. The model isn't comparing spelling; it's comparing learned meaning.
Where you'll encounter this term
- Search and recommendations: "semantic search" tools compare embeddings instead of exact keywords, so a search for "affordable laptop" can also surface a result titled "budget-friendly notebook."
- Retrieval-Augmented Generation (RAG): a common pattern where a system converts your question into an embedding, finds the most similar chunks of a document collection, and hands those chunks to an LLM as extra context before it answers.
Where you meet embeddings without noticing
Every time an AI product seems to understand what you meant rather than what you typed, embeddings are usually the reason. A support site that surfaces the right help article for a clumsily-worded question, a music service grouping songs that feel alike, a photo app finding "beach pictures" you never labeled — each works by comparing embeddings, not keywords. They are also the reason AI search can be confidently wrong: two things can sit close together in embedding space because they are usually related, even when your specific case is the exception.
Go deeper: Embeddings in the glossary · What is RAG? · NotebookLM, which searches your own documents this way
Key takeaway: embeddings turn meaning into numbers, which lets computers measure "how similar" two pieces of content are — even if they don't share any exact words.