What are embeddings? Explained with map coordinates for meaning
An embedding is a list of numbers an AI model gives to text or images so that similar meanings sit close together. Explained with map coordinates, where nearby places share a meaning.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
In one sentence: An embedding is a long list of numbers that an AI model gives a piece of text, an image or other content, like map coordinates for its meaning, so that things with similar meanings end up close together and a computer can find them by measuring distance.
The everyday picture
Think of a city map. Every building has coordinates, a pair of numbers that says exactly where it is. Two places with close coordinates are near each other, even if their names look nothing alike.
Now imagine a map where the places are not buildings but meanings. "Refund my order" and "I want my money back" share almost no words, yet on this map they sit on the same street. "Pasta recipe" is in a different part of town altogether.
An embedding is a set of coordinates on that map of meaning. Once every sentence has its coordinates, finding related ideas is simple: look at what is nearby.
The embedding model gives every piece of content coordinates on a map of meaning. A new question gets coordinates too, and the closest stored items are the most related.
How it works, step by step
- The content (a place you want on the map) can be a question, a paragraph from a manual, a product description or, with some newer models, an image or a short video.
- The embedding model (the map maker) is an AI model trained for one job: reading content and deciding where it belongs on the map of meaning. It does not write answers. It only places things.
- The embedding (the coordinates) is the list of numbers it returns. A city map needs two numbers. A map of meaning needs far more: OpenAI's text-embedding-3-small gives 1,536 numbers by default, and Google's Gemini embedding models give 3,072 by default.
- The index (the finished map) stores the coordinates of everything you have embedded, such as every page of your help centre. This is often kept in a vector database, a store built to search coordinates quickly.
- The nearby search (looking around a spot on the map) happens when a new question arrives. It is embedded the same way, and the system picks the stored items closest to it. OpenAI's guide says small distances suggest the two items are closely related.
Try it · Everyday ↔ real
Match the map to the real thing
Switch between the two and see how each part lines up.
- A place to put on the map
- The map maker
- The coordinates
- The finished map
- Looking around a spot
Putting places on a map of meaning
| Everyday | Real |
|---|---|
| A place to put on the map | Your text, image or document |
| The map maker | The embedding model |
| The coordinates | The embedding (a list of numbers) |
| The finished map | The index or vector database |
| Looking around a spot | Finding the nearest embeddings |
Where the picture stops
- A real map has two or three directions you can name: north, east, up. An embedding has hundreds or thousands, and no single number means anything a person can read. Only the overall position matters.
- Every city uses the same map. Embedding models do not. Google's documentation says embeddings from its gemini-embedding-001 and gemini-embedding-2 models cannot be compared, and that switching means re-embedding all your data. Keep one map maker for one map.
- Being close on the map means "similar in meaning", not "true" or "the right answer". A wrong page can sit very close to a question. Something still has to check what the nearby results actually say.
Try it · Myth or fact
Embeddings: myth or fact?
Decide for yourself first, then tap each statement to check it.
0 of 4 checked
Why it matters for your business
- Search that understands customers: people can find the right help article or product even when they use different words from the ones on the page. Both OpenAI and Google list search as a main use.
- Better AI answers from your own documents: embeddings are how RAG systems find the right pages to hand to an AI model before it answers.
- Sorting and spotting patterns: the same coordinates can group similar support tickets, suggest related items or flag something unusual. The providers' guides list clustering, classification, recommendations and anomaly detection.
- Cost and lock-in: providers such as OpenAI charge by the amount of text you embed, and longer lists of numbers take more storage. Changing to a model with a different map can mean re-embedding everything, so choose with care.
An example
Imagine a 25-person furniture shop with 400 help articles. Customers type "the sofa came broken" into the search box, but the right article is titled "Reporting damaged deliveries", so plain keyword search finds nothing. The shop embeds every article once and stores the coordinates. Now each search is embedded too, and the box shows the articles that sit nearest on the map. "Came broken" lands right next to "damaged deliveries", so the customer finds the right page, and the shop's team checks that the top results make sense.
Frequently asked questions
Are embeddings the same as an AI chatbot?
No. A chatbot writes text. An embedding model only turns content into numbers that describe its meaning. The two are often used together, for example in RAG.
What is the difference between embeddings and keyword search?
Keyword search matches the exact words. Embedding search matches meaning, so it can connect "money back" with "refund". Many systems use both.
Do embeddings only work for text?
No. Some models also embed images, audio, video and PDFs. Google says its gemini-embedding-2 model accepts text, images, video, audio and PDFs.
Can I mix embeddings from different models?
Not safely. Each model builds its own map, so coordinates from one cannot be compared with another. If you change models, embed your content again.
Related explainers
The DevDrift builds AI search and RAG systems that find the right page and show where each answer came from: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.