Embedding

A dense numerical vector that represents a piece of text, image, or other input in a learned space where geometric distance reflects semantic similarity.

Embeddings convert unstructured data into fixed-size vectors (typically 384–1536 dimensions) using a neural model trained on a similarity objective. Two pieces of text with similar meaning end up at nearby points; unrelated text lands far apart. The vector space is the medium for semantic search, recommendation, deduplication, and clustering.

For ecommerce, you typically embed: product titles + descriptions + attributes (one vector per item), category names, search queries, and sometimes images. Same model for queries and products is the rule — if you mix models, distance becomes meaningless.

Operational headache: when you upgrade your embedding model, the entire catalog must be re-embedded before queries with the new model will work. Many teams version their indexes and keep both running during cutover.

Related terms

Vector Search

Approximate nearest-neighbor (ANN) retrieval over high-dimensional embeddings — the engine behind semantic search, recommendation, and visual similarity.

Cosine Similarity

A distance metric that measures the angle between two vectors, ignoring their magnitude — the default similarity score for embedding-based search.

Semantic Search

Vector-based retrieval that matches by meaning rather than exact keywords, using embeddings to find products even when query and title share no words.

Cross-Encoder

A transformer that scores query and document together in a single pass — slower than bi-encoders but much more accurate, used as a re-ranker on the top-N candidates.

Browse all glossary terms