What is an embedding in an LLM?

An embedding is a way of turning text into a list of numbers that captures its meaning, so a computer can compare ideas mathematically. Instead of treating car and automobile as unrelated strings of letters, an embedding places them close together in a high-dimensional space because they mean nearly the same thing.

How does an embedding capture meaning?

Each piece of text — a word, a sentence, or a whole document — becomes a vector, essentially a long array of numbers. The trick is that text with similar meaning ends up with similar vectors, even when the exact words differ. Closeness in that space stands in for closeness in meaning, so a computer can measure how related two ideas are just by comparing their numbers.

The model that produces these vectors learned the pattern from huge amounts of text, picking up which words and phrases tend to show up in similar contexts. It never gets told that car and automobile are synonyms; it infers it from how the words are used.

What are embeddings used for?

That single trick powers a large share of modern AI. A few of the most common uses:

• Semantic search — finding results by meaning rather than exact keyword matching.
• Recommendations — spotting items similar to ones you already liked.
• Retrieval-augmented generation (RAG) — finding the most relevant documents to hand an LLM before it answers.
• Clustering and deduplication — grouping related content or catching near-duplicates.

In each case the software isn't reading the words the way you do — it's comparing vectors and trusting that nearby numbers mean nearby ideas.

Do you ever see embeddings directly?

Usually not — they work behind the scenes as the bridge between messy human language and the precise math a model can work with. A raw embedding is just a row of numbers that means nothing to a person, so you feel its effect rather than seeing it. Whenever a tool understands what you meant instead of just matching the exact words you typed — a search that finds the right answer even when your phrasing is off — embeddings are almost always doing the heavy lifting.