Embedding model
A model specialised in producing embeddings rather than generating text. Chosen on retrieval quality (MTEB score), vector dimensions, max input length and price.
What it is
Options range from hosted APIs (OpenAI text-embedding-3, Voyage, Cohere, Google) to open models you can run yourself (BGE, E5, Nomic). They differ in output dimensions (which drives storage cost), context length, and benchmark retrieval quality.
Why it matters
A higher-dimension model can retrieve better but costs more to store and search - a 3,072-d vector is 4x the storage of a 768-d one. Many models support shortening the vector (Matryoshka) to trade a little quality for a lot of cost.