Embeddings
Turning words and ideas into numbers that capture meaning.
Embeddings convert text — words, sentences, documents — into dense numeric vectors (lists of 768–3072 numbers) where semantic similarity maps to geometric closeness. "King" and "Queen" end up nearby in this space; "King" and "Invoice" are far apart.
This enables powerful operations: search by meaning (not keyword), cluster similar documents, detect anomalies, and power the retrieval step in RAG systems.
Embedding models (e.g. OpenAI text-embedding-3-small, Cohere embed-v3) are separate from generative models — smaller, faster, and cheap to run at scale.
In plain terms
Like placing concepts on a map — related ideas live close together, unrelated ones are distant. Ask 'what's nearest to Paris?' and find London, Berlin, Rome.