Vector embeddings.

What are vector embeddings in AI?

A vector embedding is a fixed-length array of floating-point numbers that represents the semantic meaning of a piece of text, image, or audio in a high-dimensional space. Items with similar meaning land close together in that space; dissimilar items land far apart. Embeddings are produced by dedicated models — OpenAI text-embedding-3-large (3072 dims), Cohere Embed v3 (1024 dims), Voyage AI voyage-3 (1024 dims), Google text-embedding-005, open-source BGE and E5 — that have been trained on hundreds of millions of text pairs. Embeddings are the foundation of semantic search, retrieval-augmented generation (RAG), clustering, recommendation, anomaly detection, and AI memory systems. Vector databases like Pinecone, Weaviate, Qdrant, Chroma, and pgvector store and index embeddings for fast nearest-neighbor lookup. Desktop assistants like Jarvis (getjarvis.eu) use embeddings to retrieve relevant memory fragments before each prompt to frontier models from Anthropic, OpenAI, or Google (with OpenAI realtime for voice). Scroll down for the embedding model comparison.