AI glossary · Building and connecting

Embedding

Also called: Text embedding · Vector

What is an embedding?

A computer normally compares words letter by letter. To it, “heating broken” and “no warmth in the flat” have nothing in common. A person sees at once that it is the same problem.

An embedding gives the computer that view. A separate AI model turns every text into a long list of numbers. The numbers stand for the meaning, not the letters. Two texts that mean similar things end up with similar numbers.

How does it work?

Picture a supermarket where goods are shelved by use instead of alphabetically. Pasta sits next to tomato sauce and parmesan, not next to nuts. Standing at the pasta, you find what goes with it right beside it.

Embeddings sort text the same way:

  • Once, in advance: Every passage in your documents is turned into numbers and gets its place on the shelf.
  • For every question: The question is turned into numbers too. The system fetches the passages that sit closest on the shelf.

Why does it matter for your business?

Embeddings are the search part of RAG, an AI that looks things up in your documents before it answers. If the search finds the wrong passages, even the best language model answers wrongly.

For your business this means: a customer writes “it stays cold in here”, and the search still finds the entry “heating broken” in the file. Nobody has to guess the exact words in the document. Anthropic also names recommendations and spotting anomalies as uses.

How does OsAI use it?

You never see embeddings, they work in the background. When OsAI builds a RAG system, for example for questions to a project file like the live demo on how it works, part of the job is making sure the search finds the right passages.

Data protection works as with any AI: if a cloud provider turns your texts into numbers, the texts go to that provider. In client projects OsAI prefers EU hosting or locally run models.

Frequently asked questions

Is an embedding the same as a language model?

No. An embedding model turns text into numbers; it does not write answers. The language model writes the answer from the passages the search found. In a RAG system both work together.

Does Anthropic offer its own embeddings for Claude?

No, according to its documentation Anthropic has no embedding model of its own (as of 13 November 2026). It names Voyage AI, owned by MongoDB, as one provider and advises comparing several.

Does the AI learn my data through embeddings?

No. Embeddings do not change the model; they only make your documents searchable. Training the model itself further is called fine-tuning, which is a different approach. If you change a document, only its embedding is recalculated.

Sources

Last updated: