Embeddings: turning meaning into numbers
An embedding is a list of numbers, a vector, that represents the meaning of a piece of text, an image or a sound. Content with similar meaning gets similar vectors, which lets software search “by idea” rather than by exact words.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|
Meaning turned into coordinates
Imagine a vast map where every sentence has a location. “How do I cancel my subscription?” and “I want to end my contract” would sit right next to each other, even though they share almost no words. “Pancake recipe” would be far away. An embedding is that location, written as a list of numbers. OpenAI’s documentation puts it simply: an embedding is a vector of numbers, and the distance between two vectors measures how related the texts are.
These vectors are computed by a neural network trained to place similar things close together. The idea was popularized in 2013 by Tomas Mikolov and colleagues at Google with word2vec, which produced one vector per word. Today’s models encode sentences, paragraphs and even images.
A concrete example
A recipe site wants a search for “quick dinner, no oven” to also surface “15-minute vegetable stir-fry”. Keyword search fails: no words match. With embeddings, the site computes a vector for each recipe once, then one for the query, and shows the closest recipes. That is the core of semantic search.
Why it matters
Embeddings are the invisible building block behind many features: document search, recommendations, duplicate detection, automatic tagging. Above all, they sit at the heart of RAG: your documents are converted into vectors and stored in a vector database, which finds the useful passages before the model answers. Embedding providers usually charge by the number of tokens processed; our token estimator helps you gauge the size of a document collection.
Common myths and pitfalls
- “Close in meaning means true.” An embedding measures similarity, not accuracy: a false but well-phrased text can sit very close to your question.
- “More dimensions is always better.” Longer vectors cost more to store and compare; the right choice depends on the task.
- “You can mix models.” Queries and documents must be encoded with the same model, or the distances are meaningless.
- “An embedding can’t reveal the original text.” That isn’t its purpose, but research published in 2023 showed that much of the original text can be reconstructed from embeddings. Treat them as sensitive data if your texts are.
Related terms: semantic search, vector database, RAG, tokenizer.
Related terms
- Vector database · A database that stores embeddings and quickly finds the ones closest to a query, which means the most relevant pieces of content.
- Semantic search · Search that finds content by meaning rather than exact words, by comparing numerical representations of the texts.
- RAG (retrieval-augmented generation) · A technique that first retrieves relevant passages from your documents, then hands them to the model so its answer is based on them.
- Tokenizer · The program that splits text into tokens (word pieces) and turns them into numbers, the only form a language model can actually read.
- Neural network · Software made of many small connected computing units that learns by adjusting the strength of the links between them.
Try it
Full glossary
Frequently asked questions
What does an embedding actually look like?
It is a long list of decimal numbers, often several hundred or a few thousand. Individually, those numbers mean nothing readable: what matters is the distance between two lists.
Do assistants like Claude or ChatGPT produce embeddings?
Embeddings come from dedicated embedding models, separate from the models that write answers. OpenAI offers embedding models in its API; Anthropic's documentation states that it does not offer its own embedding model and points to third-party providers such as Voyage AI.
Can I compare embeddings from two different models?
No. Each embedding model has its own numeric “space”. If you switch models, you need to recompute the embeddings for all your documents.
Is an embedding the same thing as a token?
No. A token is a chunk of text produced by the tokenizer. An embedding is a numeric representation of meaning, whether of a token, a sentence or a whole document.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.