Embedding (vector representation)

In short: An embedding is a list of numbers (a vector) that represents the meaning of a text, image or sound. Similar things lie close together in number space.

In more detail: “Dog” and “puppy” have similar vectors, “dog” and “tax return” do not. With embeddings you can search by meaning, not just by words: The distance (usually cosine similarity) between two vectors says how similar they are.

In Depth

• Dimensions: typically 384 to 3,072 numbers per text.
• Vector database: stores embeddings and finds the most similar quickly (for example pgvector, Qdrant).
• Use: semantic search, recommendations, duplicate detection, RAG.
• Limits: Embeddings reflect the training data. Technical terms and new words are captured less well.

See also: RAG (retrieval-augmented generation), Token (AI)