Latent Space
A latent space is the continuous representation space where a model encodes inputs. Geometry in this space (clusters, directions) often corresponds to semantic factors.
Embeddings for retrieval live in a latent space trained so that similarity matches task notions of relatedness.