What is Embedding?

Embeddings are produced by a trained model that maps raw input into a fixed-length list of numbers, learned so that semantic similarity translates into geometric proximity: 'dog' and 'puppy' land near each other, 'dog' and 'refrigerator' land far apart, even though nothing in the raw text says so explicitly. This geometric property is what makes embeddings useful for search and retrieval — you compute the embedding of a query and find the nearest stored embeddings, rather than requiring an exact keyword match. This underpins retrieval-augmented generation: instead of an LLM trying to hold all of a company's knowledge in its training data, relevant documents are found via embedding similarity search and injected into the prompt at query time, letting the model answer accurately about content it was never trained on. Embeddings are typically stored and searched in a vector database optimized for fast nearest-neighbor lookup across millions of vectors, since brute-force comparison doesn't scale.