Glossarytaught in AI0.04, AI0.01
Text embedding
Also called embeddings, vector embedding, vector embeddings, embedding model, embedding models.
A list of numbers, a vector, that a model produces to represent the meaning of a piece of text, so that passages with similar meaning sit close together. Search built on embeddings finds the paragraph about funding risk even when it never uses the words in the question.
Closeness is usually measured by cosine similarity, the product of two vectors divided by the product of their lengths. Two tiny vectors, (1, 2) and (2, 3), have a cosine similarity of 8 divided by the square root of 5 times the square root of 13, about 0.992, so they point in almost the same direction; real embeddings have hundreds or thousands of dimensions.
For an analyst the use is retrieval. A system that embeds every page of a set of annual reports can pull the passages most relevant to a question and hand them to a language model, which then answers from the documents themselves. Any answer is only as good as the passages retrieved, so check what was pulled before trusting what was said.
Similar meaning, nearby numbers.
Advertisement