AI engineering · 21 of 42

Embeddings

Meaning, written down as coordinates

Scroll

Meaning, written down as coordinates

An embedding turns a piece of text into a list of numbers positioned so that things meaning similar things end up near each other. That is the whole trick, and everything called “semantic search” rests on it.

Because position encodes meaning rather than spelling, refund policy and money back rules land close together despite sharing no words — which is exactly what keyword search cannot do.

The same property is the weakness. A typo in an error code barely moves the point, so an embedding cannot reliably tell ERR_4021 from ERR_4201. Identifiers, version numbers and part codes are where embeddings quietly fail, and why concept 8 exists.

Retrieval
MEANING, WRITTEN DOWN AS COORDINATES refund policy money back rules shipping delays embedder one point per piece of text refund policy money back shipping close = similar meaning The same idea in different words lands in nearly the same place, which is what lets a search match meaning rather than spelling. It is also why they miss exact identifiers: a typo in an error code barely moves the point.
Three phrases becoming points. The two that mean the same thing land close together.