Embedding

What is an embedding?

An AI can't read Chinese characters or look at pictures — it only understands numbers. An embedding is what turns a piece of text, an image or a sound into a fixed-length list of numbers — a vector — that carries the "meaning" of that content.

How does it work?

Encode meaning as coordinates
Picture a high-dimensional space where every word or sentence has a coordinate. Closer in meaning, closer in space — "cat" and "dog" sit far nearer to each other than "cat" and "car."
Measure similarity by distance
Whether two sentences or two images are alike isn't about the words — it's about how close their vectors are.

What is it good for?

Semantic search
Ask "how do I refund a train ticket" and the system can surface an article titled "Changes and refunds."
Recommendations
"Similar items" and "you might also like" are computed from vector similarity.
Multimodal alignment
The same thing can be mapped to nearby vectors whether it's text or an image, enabling search-by-image and search-by-text.

How do you get embeddings?

There are dedicated embedding models: feed in some content, get back a vector. A good model pushes semantically similar content close together; a bad one muddles them. That's why picking the right embedding model decides how good your search and recommendations are.

Bottom line: an embedding translates meaning into a string of numeric coordinates, so an AI can understand similarity as distance.