What are embeddings?

An embedding model turns a piece of text into a list of numbers. Texts with similar meaning end up with vectors that are close together, which lets software measure semantic similarity instead of matching keywords. Anthropic's documentation lists search, recommendations and anomaly detection among the uses.

A typical retrieval flow, as the page shows it, has three steps. Each document is converted to a vector. The query is converted to a vector too. A nearest neighbor search then finds the document whose vector is closest. The page notes that for the Voyage AI models it covers, vectors are normalized to length 1, so dot product and cosine similarity give the same ranking, and Euclidean distance ranks identically. The same models also take an input type setting that marks text as a query or a document, which the page says can improve retrieval quality.

The page states that Anthropic does not offer its own embedding model and points to Voyage AI, while advising readers to assess several vendors. Choice factors it names include how large and domain-specific the training data is, speed in production, and options to customize on private data. The details above describe Voyage AI's models, so dimensions, limits and similarity settings can differ for other providers.