Embedding
A way to turn words into numbers that AI can understand.
The glue between meaning and numbers.
Converting meaning to numbers
An embedding model reads text and outputs numbers, thousands of them, in a pattern that captures meaning. "dog" and "puppy" are close in number-space. "dog" and "algebra" are far apart.
These numbers let computers compare meaning without understanding words.
Pick a phrase to search from
Where every phrase sits in meaning-space
Nothing here reads the words. Each phrase is just a point, and phrases with similar meaning land near each other. Distance is the only thing the search measures.
Closest matches to "fast cars"
- quick vehicleskeyword search misses it
- speedy truckskeyword search misses it
- racing bikeskeyword search misses it
- linear algebrakeyword search misses it
- calculus proofskeyword search misses it
- happy dogskeyword search misses it
- playful puppieskeyword search misses it
- friendly catskeyword search misses it
The top matches share meaning. Their words can be completely different. That is why searching "quick vehicles" can surface "fast cars." A keyword search, stuck on exact words, walks right past that match.
Why they matter
Embeddings power semantic search, so you find documents about "fast cars" even when you search "quick vehicles." They also power recommendations, the kind that tell you "users who liked this also liked that."
They're also how RAG systems work. You convert a document library to embeddings, search by meaning, then hand relevant excerpts to a language model to answer.
When do you need one?
If you're building search or recommendations, yes. If you're just chatting with a model, no. Most RAG systems use a cheap standalone embedding model and save the big language model for answering.