concepts
Embeddings
Vector Embeddings
How AI represents meaning as numbers
Reading level
Embeddings are dense vector representations of text (or images, audio) in a high-dimensional space (typically 768–4096 dimensions) where semantic similarity corresponds to geometric proximity (cosine similarity). Produced by encoder models like sentence-transformers. Used in RAG for semantic search, in classification, clustering, and anomaly detection. Fine-tuning embeddings with contrastive learning (SimCSE, MNRL) improves domain-specific retrieval.
Real-world example
Spotify uses embeddings to recommend songs. Each song is embedded into a vector. When you like a song, Spotify finds other songs with similar vectors (similar musical characteristics) and recommends them.