Embeddings

Embeddings are numeric vectors — long lists of floating-point numbers — that encode the meaning of a piece of text, an image, or other data. A model trained for the job converts each input into a point in a high-dimensional space, arranging things so that inputs with similar meaning land near one another. Once meaning lives as coordinates, comparing two items becomes a distance calculation rather than a guess.

That property is what makes embeddings so useful. To search by meaning, you embed every document once, embed the incoming query, and return whichever documents sit closest — typically by cosine similarity. This is the foundation of semantic search and the retrieval step inside retrieval-augmented generation (RAG). At scale, the vectors are stored and searched in a purpose-built vector database so that nearest-neighbor lookups stay fast over millions of items. The embedding model itself is usually a specialized cousin of a large language model, tuned to output one meaningful vector per input rather than generating text.

A concrete example: embed the sentences "How do I reset my password?" and "I forgot my login credentials." The literal words barely overlap, yet their vectors sit close together, so a semantic search for one will surface the other. A plain keyword search, by contrast, would miss the match entirely — and that gap is exactly why embeddings became a core building block of modern AI systems.