Definition
An embedding is a list of numbers that represents an input, such as a sentence, image, or product description, in a learned vector space. Inputs that the embedding model maps as similar tend to have nearby vectors under a compatible similarity or distance measure.
The embedding is the representation. The embedding model produces it.
Simple example
The phrases “stop a background worker” and “cancel a hosted service” use different words but can produce nearby text embeddings. A search system can compare the vectors and find the second phrase for a query containing the first.
The individual numbers are not labels. One coordinate does not reliably mean “background work” while another means “cancellation.” The vector works as a whole.
Why it matters
Embeddings let applications compare meaning-like patterns without requiring exact keyword matches. They are used for semantic search, recommendations, clustering, deduplication, and retrieval pipelines.
Keep each embedding associated with its source and required metadata. The vector does not replace the original content, provenance metadata, or access-control data.
One important nuance
For meaningful comparisons, vectors need a compatible embedding contract. That includes the embedding model and version, dimensions, required input preparation, and similarity or distance measure. Queries and documents may use different preparation when the model expects it. Two vectors with the same length are not necessarily in the same space.