Embedding

An embedding is a technique that converts the "meaning" of words into numerical coordinates. It's like plotting each sentence on a map — sentences with similar meanings end up close together. "Recommend dog treats" and "nutritional supplements for pets" barely share any words, but in embedding coordinates, they're neighbors.

This matters because computers can finally compare things by meaning rather than by exact words. Old-school search needed matching keywords, but embedding-based search (semantic search) finds results that share meaning even when the wording is completely different.

It's the most widely used component behind the scenes in AI products. Search, recommendations ("articles similar to this one"), and the "retrieve relevant material" step in RAG all run on embeddings. When an article mentions a "vector database" or "semantic search," it's really talking about embeddings.

"They improved search accuracy with a new embedding model" — this means they made the "meaning coordinates" used for meaning-based search more precise.

1. Try sending this to a chatbot: "Group these five sentences by similar meaning: ① Recommend dog treats ② The stock market crashed ③ What are good supplements for dogs? ④ The KOSPI dropped sharply ⑤ Will it rain tomorrow?"
2. Sentences ①③ and ②④ get grouped together — even though they barely share any words. They're grouped by "distance in meaning" rather than by words, and embeddings are the coordinates used to measure that distance.
3. The "related articles" section at the bottom of METAL articles is pulled the same way — embeddings do the work of finding pieces that connect in content even when the wording differs.