Embedding model
A model that converts text (words, sentences, documents) into a vector of numbers that captures its meaning, so semantically similar texts end up close together in the vector space. The engine behind semantic search, RAG, and recommendations.
Mechanic: unlike a generative LLM (which produces text), an embedding model produces a fixed vector (often 384–3072 dimensions) per text. The vectors are constructed so that distance/cosine similarity corresponds to meaning similarity → "dog" and "puppy" lie close, "dog" and "quantum physics" far apart. Used for: vector search (find documents similar to a query), clustering, classification, deduplication, recommendations. Examples: OpenAI text-embedding-3, Cohere Embed, open source like BGE, E5, Sentence-Transformers (SBERT). A "bi-encoder" embeds the query and document separately (fast, searchable), unlike a reranker/cross-encoder (compares a pair at a time, more accurate but slower). The base component of all vector-based AI. Related to matryoshka embedding and embedding drift.