What is an embedding?

Computers cannot compare meaning directly, so AI systems convert content into embeddings: long lists of numbers produced by a model, where pieces of content with similar meaning end up with similar numbers. A sentence about a late delivery and one about a package that never arrived would have close embeddings, even though they share few words.

This is what makes searching by meaning possible. A company might create embeddings for every article in its help center. When a customer types a question, the question is turned into an embedding too, and the system finds the articles whose embeddings are closest, returning the right billing article even if the customer used different words from the article.

Embeddings are usually stored in a vector database, a database designed to find the closest matches among very large numbers of these lists quickly. They power the search step in retrieval-augmented generation, and they are also used to group similar feedback, spot duplicate support tickets, and recommend related content.

Embeddings work behind the scenes, so most people never see them directly. They are not perfect: two passages can look similar in meaning to the model while differing in a detail that matters, such as a date, an amount, or whether something is allowed or forbidden. Search results based on embeddings still need checking, especially when exact wording carries weight, as in contracts.

An example

An internal search tool uses embeddings so that a search for time off finds the leave policy, even though the policy never uses that phrase.