If you have read three blog posts and still aren't sure what a vector is doing, this one is for you. We will skip the math we don't need and keep the math we do.
What an embedding is
An embedding is a list of numbers — typically 1024 of them — that summarizes a piece of text. The numbers are not assigned by hand; they are produced by a model trained on a great deal of text. The model's job, during training, is to produce embeddings such that semantically-similar pieces of text end up with similar numbers.
Similarity, by example
"Cancel my subscription" and "How do I unsubscribe?" come from different words but mean roughly the same thing. Their embeddings are close together in the 1024-dimensional space the model defines. "What's the weather?" comes from words that look unrelated, and its embedding sits far away.
This is the trick that makes retrieval work. You embed the user's question, you embed every chunk in your knowledge base, and you fetch the chunks whose embeddings are closest to the question's. "Closest" is usually measured by cosine similarity — how aligned the two vectors are, regardless of magnitude.
What the dimensions mean
They don't mean anything individually. Don't go looking for the "sentiment dimension" or the "topic dimension." The space is learned and the dimensions are entangled. What matters is the relative geometry — which embeddings are close to which other embeddings.
Why dimension count matters
More dimensions means more capacity to distinguish meanings. 1024 is a common production default; some models go to 3072 or higher. More dimensions also means more storage and slower search. Pick a model whose dimension count fits your scale.
What embeddings cannot do
- They don't reason. Two sentences can mean opposite things and still be embedded close together if they share enough vocabulary.
- They don't handle negation well. "This is great" and "This is not great" are closer than they should be.
- They are domain-shifted. A general-purpose embedding model may not understand your jargon. Fine-tuned or domain-specific models often help.
- They are not search. They are one ingredient. The full retrieval pipeline (parse, chunk, enrich, hybrid retrieve, rerank) does the work.
The smallest possible mental model
Embeddings let you compare meanings using arithmetic. That's the whole game. Everything in retrieval is what you do with that ability.