← AI Terminology

Embedding

An embedding is a dense, low-dimensional vector representation of a discrete object (word, sentence, image, user, product) that captures semantic similarity — similar things map to nearby vectors in the embedding space.

Embeddings are the bridge between discrete real-world objects and the continuous vector operations that neural networks perform.
Why It Matters in AI
Before embeddings, words were one-hot vectors (sparse, no notion of similarity). Word2Vec showed that "king − man + woman ≈ queen" — meaning is encoded in vector arithmetic. Today, sentence and document embeddings power semantic search, RAG, recommendation systems, and clustering. LLM token embeddings are the input representation that enables language models to process text at all.
Key Points
Aspect Description
Similarity Cosine similarity or dot product between vectors → semantic distance metric
Applications Semantic search, RAG retrieval, clustering, recommendations, anomaly detection, few-shot
Sentence/doc BERT, sentence-transformers — full sentence → 768–1536 dim vector capturing semantic meaning
Word embeddings Word2Vec, GloVe, FastText — word → 100–300 dim vector trained on co-occurrence
Embedding models text-embedding-3-large (OpenAI), E5, BGE, Nomic Embed, Cohere Embed
Image embeddings CLIP, ViT — image → vector in same space as text embeddings for cross-modal search
Simple Analogy
A map coordinates system for meaning: London (51.5°N, 0.1°W) and Paris (48.9°N, 2.3°E) are geographically close; Tokyo (35.7°N, 139.7°E) is far. Embeddings create a similar coordinate system for concepts: "cat" and "kitten" are nearby; "cat" and "democracy" are far apart. Cosine similarity is the distance measurement.
Common Usage Examples
  • openai.embeddings.create(model="text-embedding-3-large", input="...") — 3072-dim vectors
  • sentence_transformers.SentenceTransformer('all-MiniLM-L6-v2').encode(texts) — fast sentence embeddings
  • Semantic search: embed query + documents, return top-k by cosine similarity
  • RAG: chunk documents → embed → store in vector DB (Pinecone, Weaviate, Chroma) → retrieve on query
  • Product recommendations: embed user history + items → find nearest product embeddings
Summary
In short: Embeddings turn discrete objects into vectors where distance = similarity — the universal representation layer enabling semantic search, RAG, and every other similarity-based AI application.