An embedding is a numerical representation of data, usually a list of numbers called a vector, that captures meaning so similar things end up close together. A sentence about refund policies and a question about getting money back map to nearby vectors even though they share few words. Text, images, audio, and table rows can all be embedded.
Embeddings power semantic search, recommendations, clustering, and retrieval augmented generation, where a system finds relevant documents by comparing vectors rather than matching keywords. Models learn embeddings during training, and dedicated embedding models turn any input into vectors on demand.
Embeddings inherit everything in the data they were trained on and applied to. Outdated or inconsistent source records produce vectors that quietly point retrieval in the wrong direction, which is why embedding pipelines need the same versioning discipline as any other data pipeline.
Related terms: Vector Database · Retrieval-Augmented Generation (RAG) · Unstructured Data · Semantic Layer