Embeddings
Dense numerical vector representations of text, images, or audio — used to measure semantic similarity, power search, and ground LLM outputs in your data.
Embeddings turn content into vectors so machines can reason about meaning. The same idea expressed in different words produces similar vectors; unrelated topics produce distant vectors. This single property powers semantic search, RAG, recommendation systems, clustering, and anomaly detection.
In 2026, the practical question is which embedding model to use. OpenAI text-embedding-3 family for English, Cohere Embed v3 for multilingual, Voyage AI for RAG-tuned, and BGE / E5 for self-hosted. Pick by language coverage, task type, and cost.
For agent builders, embeddings are the unsung workhorses. Document chunking, query similarity, memory retrieval — all run on embedding vectors stored in [vector databases](/glossary/vector-database). Embedding quality bounds the entire system's output quality.
Frequently asked
How do embeddings differ from one-hot encoding?+
One-hot encoding treats each word as completely distinct. Embeddings place semantically similar words near each other in vector space. The latter captures meaning; the former does not.
What is the right embedding dimension?+
Common defaults: 768 (smaller, faster), 1536 (OpenAI ada-002), 3072 (OpenAI text-embedding-3-large). Higher dimensions improve accuracy marginally but multiply storage and search cost. Most production stacks use 1024–1536.