What are embeddings?

At a glance

• Computers match by meaning, not keywords: ‘cancel my plan’ finds an article titled ‘ending your subscription.’^[1]
• Closeness equals similarity. Every item is a point; the system answers by finding the nearest ones.^[2]
• They power semantic search, recommendations, and ‘chat with your documents’ AI (RAG).
• You buy embeddings, not build them: call a model, store results in a vector database.

How it works

An embedding model gives each piece of content coordinates on a map of meaning. Because meaning becomes distance, the computer answers ‘what is this most like?’ by finding the nearest points. The classic proof: the math ‘king minus man plus woman’ lands near ‘queen.‘^[3]

What it powers

Semantic search finds results by intent, tolerating typos and slang. Recommendations surface items nearest to what someone liked. RAG lets a chatbot answer from your own files: documents and the question both become embeddings, the closest passages are pulled, then the AI writes a grounded answer.^[4]

Before you trust it

Embeddings are cheap; the real risk is fit. A model strong on web text can be weak on your contracts or catalog, and public leaderboards are self-reported.^[5] Ask vendors which model they use, whether your data leaves your environment, and to show retrieval accuracy on a sample of your real content.

Bottom line

Embeddings turn meaning into distance so software finds what is similar, not just matching words; the decision that matters is how well a model retrieves answers on your own data.