What is an embedding

Definition

An embedding turns a piece of text, or an image, into a list of numbers chosen so that similar meanings get similar numbers. Texts that mean the same land close together even when they share no word, which is what lets an AI tool find the passage a customer needs. Real embeddings have hundreds to thousands of numbers: OpenAI's text-embedding-3-small returns 1,536 by default and text-embedding-3-large 3,072. The maps that show embeddings as dots on a page are simplified to two dimensions for teaching. Embeddings are the first step of a vector database and of RAG.

Why it matters

An owner rarely creates them by hand, but they sit under every tool that answers from your files, so two facts matter: they match meaning, not words, and they carry the content of the text, so they need the same care as the documents.

Directly related: Vector Database, RAG, Token.

Improvement tips

• Use one embedding model for everything you compare, because numbers from two different models do not match.
• Keep passages short, one rule or topic each, so each embedding stands for one meaning.
• Test with questions in your customers' own words and in the language they write in.
• Protect stored embeddings like the documents they came from.

Common mistakes

• Mixing embeddings from two models, so the search compares numbers that do not match.
• Embedding long pages that mix many topics, so no single passage matches a question well.
• Expecting embeddings to judge whether a passage is true or current, when they only measure closeness of meaning.