Embeddings

Embeddings are dense vector representations of text, images, or other objects that place similar items near each other in vector space.

They turn meaning into coordinates so “nearby” means “related.”

Embedding models map objects to vectors such that semantic similarity correlates with geometric proximity (often cosine or dot-product). They power dense retrieval, clustering, and recommender features.

Embeddings are the shared currency between search, RAG, and many personalization systems.

• Semantic search and RAG
• Duplicate detection
• Recommendation candidate generation

• Meaning-aware similarity
• Language-flexible matching
• Composable building block across products

• Domain mismatch hurts quality
• Vector similarity is not logical entailment
• Requires careful evaluation of retrieval metrics

• Higher dimensions are not automatically better
• Embeddings do not store full documents by themselves

Do embeddings replace keywords?

Often they complement keywords in hybrid search rather than fully replacing lexical methods.

• NIST AI RMF (risk context for AI components)

• Mikolov et al. — word2vec (historical embedding methods)