Embedding
An embedding represents text, an image or audio as a list of numbers in which things close in meaning end up close together. Semantic search, recommendations and the whole RAG pattern rest on it.
How it works
A model turns a fragment into a vector, a list of hundreds or thousands of numbers. No single number means anything on its own, but the distance between two vectors reflects how close their meanings are.
So "car" and "automobile" land near each other while "car" and "soup" do not, even though the first pair shares no letters. Embedding search finds by meaning rather than by string match.
The same embedding model must be used for indexing and for querying. Vectors from different models are not comparable: change the model and the store has to be rebuilt.
Where it is used
- Semantic searchFinds the right document from a description of a problem, even when the user does not know the correct terms.
- DeduplicationTwo texts about the same thing sit close together in vector space even when written in different words.
- RecommendationsSimilar articles, products or tracks are matched by vector proximity rather than by shared tags.
- ClassificationTopic or sentiment can be decided by which known examples a vector lands nearest to.