Embeddings

An embedding is a list of numbers, typically a few hundred to a few thousand, produced by a model that has read a piece of text. The numbers carry no meaning alone; the position they define does. Text about invoices is placed near other text about invoices and far from text about tennis, because the model was trained to put related meanings close together.

Position is what makes semantic search work. Once meaning is a position, similarity is distance, and distance is arithmetic a computer can do across millions of items in milliseconds. Embed the question, find the nearest stored chunks, hand them to the model. No keyword had to match.

The coordinates only mean something relative to the model that produced them. Query with a new embedding model against chunks stored by an old one and every comparison is nonsense, returned as a confident, well-ordered list of the wrong chunks. Changing the embedding model means re-indexing everything, not editing a config value.