Back to the index
13/ 55

HOW MODELS WORK

Embedding.

Vector representation

A numerical representation that places items in a space where useful relationships can be compared.

In plain words

An embedding represents an item, such as a passage, image, or token, as a vector: an ordered list of numbers. Learning shapes this space so that useful relationships can be captured mathematically.

A closer look

For semantic search, an embedding model maps both a query and candidate documents into compatible vectors. A similarity measure then helps rank candidates. This can find related ideas even when they use different words. Other embeddings are learned inside models to represent tokens or intermediate features.

What counts as “similar” depends on the model and its training objective. Two texts may share a topic while disagreeing on a fact. Vector distance is therefore a retrieval signal, not a truth test. Representations from different models are generally not directly comparable without an explicit mapping.

In practice

AN EXAMPLE

A search for “how to reset my password” might retrieve a page titled “Recover access to your account,” because their embeddings capture a useful semantic relationship.

A useful distinction

An embedding is not a readable dictionary of an item’s meaning. Individual dimensions usually do not correspond to tidy concepts, and numerical proximity does not prove equivalence, relevance, or correctness.

Watch & learn

Sources & further reading

Google — Embeddings (opens in a new tab)