← AI Terminology

Embedding Model

An embedding model maps text, images, or other inputs into fixed-size vectors that place similar items close together for search, clustering, and retrieval.

Sentence transformers and vendor embedding APIs are the workhorses of RAG.
Why It Matters in AI
Every semantic search stack is only as good as its embedding model. Domain mismatch (legal, code, medical) needs the right embedder. Choosing, evaluating, and versioning embedding models is foundational ML product work.
Key Points
Aspect Description
APIs OpenAI, Cohere, Voyage; open: E5, BGE, GTE
Dims Typically 384–3072; affects storage and speed
Eval MTEB leaderboard, domain recall@k
Types Text, multimodal, code-specialised embedders
Related Vector DB, cosine similarity, dense retrieval
Training Contrastive learning on pairs/triplets
Simple Analogy
A translator that turns any sentence into GPS coordinates on a meaning map — similar sentences land in nearby neighbourhoods.
Common Usage Examples
  • SentenceTransformer('all-mpnet-base-v2')
  • OpenAI text-embedding-3-large
  • MTEB to pick models
  • Re-embed corpus when changing models
Summary
In short: An embedding model converts inputs into vectors for similarity — the encoder that makes semantic search and RAG possible.