How do embeddings work?
An embedding model is trained so that texts with similar meanings end up close together in a space with hundreds or thousands of dimensions. OpenAI’s embeddings guide, checked on 8 October 2026, describes an embedding as a vector of floating-point numbers where the distance between two vectors measures their relatedness: small distances suggest high relatedness. Its text-embedding-3-small model returns 1,536 numbers by default, and text-embedding-3-large returns 3,072.
Systems usually compare embeddings with cosine similarity, which OpenAI recommends. Because OpenAI’s embeddings are normalised to length 1, that is the same as a dot product. Google’s machine learning glossary adds that the values change every time a model is retrained, so embeddings from different models, or different versions of one, can’t be compared directly.
In a support knowledge base, “How do I add a teammate?” and “Inviting colleagues to your workspace” share no important words, yet an embedding model places them close together, so a semantic search can match one to the other.
Why do embeddings matter for a voice agent?
Embeddings are the usual way an AI agent finds the right passage when a caller’s words differ from the docs. A voice agent built on a typical RAG pipeline embeds its knowledge base in advance, embeds each search query during the call and fetches the closest passages.
They come with upkeep. Every document has to be re-embedded when it changes, the vectors have to be stored somewhere, and results depend on how documents are split into chunks. For a small, curated set of documents, some products use keyword search instead.
Does Kawl Now use embeddings?
No. Kawl Now doesn’t create or store embeddings of your knowledge. Approved sources are kept as plain text in Postgres, and the agent’s search matches words rather than vectors, so there is no embedding model to choose and no index to rebuild when you edit a source.
The language model behind the agent uses embeddings internally, as Transformer models do, but your sources aren’t sent to an embeddings service. They reach the voice service only as search excerpts during a call.
An edited source is searchable as soon as it’s approved. Because nothing bridges synonyms for you, write sources in the words your customers use.
Common questions
What is the difference between an embedding and a vector?
A vector is any ordered list of numbers. An embedding is a vector produced by a model to represent a piece of data, arranged so that similar items get similar vectors.
Are embeddings the same as tokens?
No. Tokens are the pieces text is split into, such as words or parts of words. An embedding is the list of numbers a model uses to represent a token, a sentence or a whole document.