Technique
What is Embedding?
Also known as: vector embedding
An embedding is a list of numbers representing a piece of text, image or audio in a way that places similar meanings close together — the numeric form that makes semantic search possible.
A model converts cancel my subscription into a vector of maybe 1,536 numbers — coordinates in a high-dimensional space arranged so that meaning becomes geometry. End my plan lands nearby despite sharing no words; renew my plan lands further away despite sharing two. Because closeness is now arithmetic, comparing meaning becomes a distance calculation you can run over millions of items in milliseconds. Embeddings come from their own models, separate from the chat model, and vectors from different embedding models are not comparable.
Embeddings are what make retrieval work — RAG over your documents, and tool discovery over a large catalog. In Velaris, connectors are data in a searchable capability mesh: the agent describes what it needs in plain terms and the mesh returns the closest-matching capabilities, which is why a thousand tools don't have to be a thousand schemas in the prompt. The caveat worth knowing: embeddings capture similarity, not truth or intent. Delete the file and restore the file sit close together, so retrieval finds candidates — the agent still has to choose.