Embedding

A list of numbers (a vector) that represents the meaning of a piece of text, image, or audio so that similar things sit near each other in vector space.

An embedding is a fixed-length list of numbers — typically 384, 768, 1536, or 3072 dimensions — that encodes the meaning of an input. "Cat" and "kitten" produce nearby vectors; "cat" and "calculus" sit far apart. The model that produces embeddings is trained so that semantic similarity becomes geometric closeness. They matter because embeddings are how machines do meaning-based search. Once you've turned every help-center article, product, image, or chunk of code into a vector, you can find the most relevant ones for a query by computing cosine similarity in milliseconds — even across millions of items. That's what powers RAG, semantic search, recommendation, deduplication, and clustering. A concrete example: you have 50,000 customer support tickets. Embed each one (OpenAI's text-embedding-3-small or BGE), store the vectors in a vector database, and now "my password reset email never arrived" can find tickets about "verification email missing" without sharing any keywords. Pre-embedding, you'd have needed exact-match search or a labeled classifier. Embeddings are also how you connect modalities — CLIP embeds images and text into the same space, so "a red sports car" can find matching photos. Related: vector database, RAG, semantic search, cosine similarity, CLIP.