← Explainers
2 min read

What are embeddings?

How computers turn meaning into numbers, so they can tell that 'refund' and 'money back' mean the same thing.

BuzzwordsAI basics

Search for “money back” and a good help centre still finds the article titled “Refund policy”, even though the words don’t match. The trick behind that is embeddings.

Term
Embeddings
Meaning, turned into numbers

A way of turning text into a list of numbers that captures its meaning, so a computer can measure how similar two pieces of text are.

Think of it like this

A map of ideas. Every word or sentence gets a location. Things with similar meanings sit close together: “puppy” near “dog”, “refund” near “money back”, and both far away from “password reset”.

How it works

From text to meaning
  1. 01Text
  2. 02Embedding model
  3. 03List of numbers
  4. 04Compare distances
An embedding (shortened)
0.12-0.480.330.91…0.07
Real ones have hundreds or thousands of numbers.

Those numbers are coordinates. Plot them and words with similar meanings land near each other.

A map of meaning (2 of the hundreds of dimensions)
Money Account Delivery refund money back chargeback password login reset shipping courier tracking
Illustrative. Real embeddings have hundreds of dimensions; this shows two.

Measuring similarity

Once text is turned into numbers, a computer can check how close two meanings are.

How close is each one to "I want my money back"?
  • Can I get a refund?Very close
  • My order arrived damagedSomewhat related
  • How do I reset my password?Not related

Where you’ll find them

  • Smarter search: finding results by meaning, not just matching words.
  • Recommendations: “people who liked this also liked…”
  • Grouping: sorting thousands of support tickets by topic automatically.
  • RAG: finding the right documents to give an AI before it answers. See RAG.

Embeddings are usually stored in a vector database, which is built to search millions of them quickly.

A common misunderstanding

“The computer understands meaning like we do.” It doesn’t. Embeddings capture patterns in how words are used together. That’s enough to be very useful, but it isn’t human understanding.

A typical case

Online shops use embeddings so a search for “running shoes” also finds products listed as “trainers”, even though the words do not match.

RAG · Token · LLM