Embeddings

How AI represents meaning as vectors in high-dimensional space.

What are Embeddings?

Embeddings are dense vector representations that capture semantic meaning. Similar concepts have similar embeddings, enabling machines to understand relationships between words, sentences, and documents.

How Embeddings Work

Embedding models map discrete tokens to continuous vectors in a high-dimensional space (often 768-4096 dimensions). The position of each vector encodes its semantic meaning.

Semantic Similarity

Similar meanings cluster together in embedding space. "King" and "Queen" are closer than "King" and "Banana".

Vector Dimensions

Each dimension captures some aspect of meaning—though these dimensions aren't human-interpretable.

Vector Operations

Famous example: King - Man + Woman ≈ Queen. Relationships are encoded as directions in the space.

How Embeddings Are Created

Embeddings come from the embedding layer—a learned lookup table that sits at the very beginning of a neural network.

Token
"cat" → ID: 2368
The Embedding Layer
matrix[2368]
Vector
[0.23, -0.87, 0.41, ...]

The Embedding Layer

When a token enters the model, its ID is used to look up a row in a large matrix. This row IS the embedding—a dense vector of learned weights.

Learning Through Training

During training, the embedding weights are adjusted via backpropagation. Words that appear in similar contexts develop similar embeddings.

In LLMs

The embedding layer converts each token ID to a vector. These vectors are then processed through transformer layers, combined with positional encodings to understand word order.

Dedicated Models

Models like text-embedding-3-small or all-MiniLM are trained specifically to produce embeddings useful for similarity search, with contrastive learning objectives.

Embedding Dimensions

Embedding size varies: GPT-2 uses 768d, GPT-4 uses 12,288d, dedicated embedding models often use 384-1536d. Larger isn't always better—it depends on the task.

Common Use Cases

Semantic Search

Find documents by meaning, not just keyword matching.

Clustering

Group similar documents, detect topics automatically.

Classification

Categorize text based on embedding similarity to examples.

RAG Systems

Retrieve relevant context for LLM prompts.

🎯

Interactive Visualization

Explore how embeddings cluster by meaning

Select words to visualize

Click words to add/remove from visualization

animals
royalty
technology
food
emotions
5 words shown

Vector Space (2D Projection)

2D projection of embedding space

+x-x+y-y0kingqueencatdogcomputer
animals
royalty
technology
food
emotions

Key Takeaways

  • 1Embeddings convert text to vectors that capture semantic meaning
  • 2Similar concepts have similar embeddings (cosine similarity)
  • 3Embeddings enable semantic search, clustering, and RAG
  • 4Different embedding models have different strengths and dimensions