What are embeddings?

Embeddings are how AI turns meaning into maths. An embedding model reads a piece of content — a word, a sentence, a product description, an image — and outputs a long list of numbers, often hundreds or thousands of them. That list is a point in a high-dimensional space. Content with similar meaning lands near each other.

How it works

The embedding model has learned, from huge amounts of data, which concepts tend to go together. So “How do I reset my password?” and “I forgot my login” produce vectors that are close, even though they share almost no words. To compare two items, software calculates the distance or angle between their vectors (cosine similarity is a common method). Closer means more related.

This is why embeddings beat keyword search for many tasks: they match intent, not spelling.

An analogy for the meatbags

Imagine a giant library where books aren’t shelved alphabetically but by what they’re about — cookbooks in one corner, Italian cookbooks in a tighter cluster within it, pasta books tighter still. An embedding is a book’s exact shelf coordinates. To find related books, you don’t read titles; you look at what’s physically nearby.

Why it matters for the Chosen

Embeddings are the quiet infrastructure behind many profitable AI products:

• Retrieval-augmented generation. RAG systems embed documents, then retrieve the closest chunks to a user’s question before the model answers.
• Semantic search on websites, help centres and internal wikis — customers find answers even when they phrase questions badly.
• Recommendations: “products like this”, “articles you might like”.
• Deduplication and clustering: grouping thousands of support tickets or survey responses by theme.
• Classification: routing emails or leads by comparing them to labelled examples.

Embeddings are usually stored and searched in a vector database. If you’re building a “chat with your documents” tool or smart search for a client, you’re working with embeddings whether the no-code platform shows them to you or not.

Things to know

• Embeddings from different models aren’t compatible; switching models means re-embedding everything.
• How you chunk documents strongly affects retrieval quality.
• Similar isn’t the same as correct — a passage can be semantically close to a question without containing the answer.
• Multimodal embedding models can place text and images in the same space, so a text query can find matching pictures.