Vector embedding
A dense numerical vector representation of text, image, or audio — produced by an embedding model and used to measure semantic similarity in high-dimensional space.
A vector embedding is what an LLM uses to think about meaning. Text goes in; a fixed-length vector (typically 768 to 3072 dimensions) comes out, where similar meanings cluster together in space. "Cat" and "kitten" end up close; "cat" and "diesel engine" end up far apart.
For agents, embeddings power retrieval at every layer. Documents in your knowledge base get embedded once and stored in a vector database. Each user query gets embedded and matched against the stored set. Memory systems embed past interactions for recall. Recommendation systems embed actions for personalization.
Frontier embedding models in 2026: OpenAI text-embedding-3-large (3072 dims, English-strong), Cohere Embed v3 (multilingual), Voyage AI (RAG-optimized), and open-source BGE/E5 families. Pick by language coverage, retrieval task type, and cost — not by claimed benchmark wins, which often do not transfer.
Frequently asked
How do I choose an embedding model?+
Start with cost and language coverage. For English-only, OpenAI text-embedding-3-small is fast and cheap. For multilingual, Cohere Embed v3 or Voyage Multilingual. For self-hosted, BGE-large or E5-mistral-7b. Validate on your own retrieval task.
Are larger embeddings always better?+
No. Higher dimensions improve retrieval accuracy marginally but quadruple storage and search cost. Most production stacks use 768–1536-dimensional embeddings as the sweet spot.