Nested Embedding
A representation technique where lower-dimensional embeddings are strictly contained as a prefix within the dimensions of a larger, full-size embedding.
Think of It Like This
Like a Russian nesting doll, where the core miniature doll perfectly fits inside and forms the center of every larger doll.
Nested embeddings allow developers to truncate a single large vector (e.g., 2048 dimensions) down to a smaller size (e.g., 256 dimensions) without drastically losing semantic integrity. This flexibility enables cost-effective storage in vector databases while preserving accuracy. It is achieved during training by explicitly applying loss to various truncations.