In one sentence
An embedding is a way to represent a piece of text as numbers so software can compare how similar two pieces are in meaning.
The simple model
For now, think of an embedding as placing each piece of text somewhere on a giant invisible map, where things with similar meaning sit close together. "Cancel my subscription" and "I want to end my membership" land near each other, even though they share almost no words.
Software cannot compare meaning directly, but it can compare positions on that map. Embeddings are what turn "how similar in meaning are these?" into a calculation a computer can do.
Why you care
Embeddings are the machinery behind semantic search, which is how a RAG system finds relevant passages even when the question uses different words from the document.
You almost never need to work with embeddings directly. What matters is understanding what they make possible: finding related information by meaning rather than exact keywords, and knowing where that can go wrong.
Where the simple picture breaks
The invisible map is a helpful mental model, not the full mechanism. A few practical consequences are worth knowing:
- Similarity is about wording and meaning, not truth. Two passages can be close on the map and still contradict each other.
- The quality depends on the model that creates the embeddings, and different models place things differently.
- Very long pieces of text embed poorly, which is why documents are split into chunks first.
A small example
A support tool needed to find past tickets similar to a new one. Keyword search missed tickets that described the same problem in different words. Embedding the tickets and searching by meaning surfaced the genuinely similar cases, which keyword matching had hidden.
What this does not solve
Embeddings help software find related text; they do not judge whether that text is correct, current or the best answer. A close match can still be the wrong source. Retrieval by meaning is a starting point, not a guarantee.
