# What is Embedding Model?

Also called embeddings, vector model.

An embedding model converts text, images, or other data into a fixed-length list of numbers, called a vector, that represents meaning. Items with similar meaning land close together in that numeric space, so similarity can be measured arithmetically. Embeddings underpin semantic search, recommendation, clustering, deduplication, and the retrieval step in retrieval-augmented generation.

Similarity is computed with a simple distance measure, most often the cosine of the angle between two vectors. Because the comparison is pure arithmetic, millions of stored items can be searched in milliseconds using a vector index. This is what allows a query about canceling a plan to retrieve a document titled 'ending a subscription' despite sharing no keywords.

Embedding models are much smaller and cheaper to run than generative ones, and they emit vectors rather than text. They cannot answer a question or write a sentence. A typical production system therefore uses both together: an embedding model finds the relevant passages, and a language model reads those passages and composes the actual response for the user.

Vectors are only comparable within a single model. Changing embedding models means re-embedding the entire corpus, since two models place the same sentence in unrelated coordinate systems with no translation between them. Dimension counts differ as well, commonly from a few hundred to a few thousand, trading storage size and search speed against representational fidelity.

Semantic similarity is not the same thing as relevance, and this trips up many teams. An embedding search will happily return a passage that is topically close but factually contradictory, because 'is covered' and 'is not covered' sit very near each other in the space. Production systems therefore add keyword matching and a separate reranking step.

## Key points

- Turns content into vectors where distance approximates similarity of meaning.
- Powers semantic search, clustering, deduplication, and retrieval pipelines.
- Smaller and cheaper than generative models; produces no text.
- Vectors from different models are incomparable, so switching means reindexing.
- Semantic closeness can retrieve contradictions and negations.

## In practice

A help center embeds every article once and stores the vectors. A user types 'my card was charged twice'. That sentence is embedded and compared against the stored vectors, and the closest match is an article titled 'Resolving duplicate payments', which shares no words with the query. That article is then handed to a language model, which writes a specific answer.

## Related terms

- [Token](/en/glossary/token)
- [Tokenization](/en/glossary/tokenization)
- [Large Language Model](/en/glossary/large-language-model)
- [Multimodal Model](/en/glossary/multimodal-model)
- [Inference](/en/glossary/inference)

[Back to the AI Glossary](/en/glossary)