Embeddings in plain English: how software finds things by meaning
Search used to mean matching words. If you typed “money back” and the policy said “refund”, you got nothing. Embeddings fix that, and they are the part of retrieval that most people find mysterious. The idea is simpler than it sounds.
Meaning as a position
An embedding model reads a piece of text and returns a long list of numbers. You can think of those numbers as a position on a very large map. Texts that mean similar things land near each other. “How do I get my money back?” and “Refund policy” end up close. “Office holiday calendar” ends up somewhere else.
Figure 1 · Closeness in meaning
How it’s used
- Turn each chunk of your documents into an embedding and store them.
- When a question comes in, turn it into an embedding too.
- Find the stored chunks closest to the question.
That is the retrieval step in RAG. The model then answers from the chunks it was handed.
What embeddings don’t do
- They measure closeness, not truth. A chunk can be close in topic and still not contain the answer.
- They can miss exact details. Part numbers, names and negations (“does not apply”) are easy to blur.
- They depend on how you cut the text. A chunk with no context around it embeds poorly.
- Jargon can trip them up. Terms specific to your company may need testing.
How to check yours
Take twenty real questions, and for each one look at what the system retrieved. If the right passage isn’t in the top few, the fix is upstream of the model: better chunks, better search, or both. It is the same habit as in the spreadsheet eval method.