What are embeddings?
Embeddings are the reason AI document search feels different from the search boxes you're used to. You ask about something in your own words, and it finds the right document even though that document uses entirely different terminology.
Meaning as coordinates
An embedding model reads a piece of text and outputs a long list of numbers — often several hundred. Those numbers are a position: a point placed in a space where distance corresponds to similarity of meaning.
"The invoice is overdue" and "payment hasn't come through" share almost no words, but land close together. "The invoice is overdue" and "the weather is fine" land far apart.
Nobody designed those coordinates by hand. They emerge from training on enormous amounts of text, where the model learns which expressions get used in similar ways.
What this makes possible
- Search that understands paraphrase. Ask about "time off" and find the leave policy.
- Grouping related material. Cluster support tickets by what they're actually about rather than by keyword.
- Recommending the relevant next thing. Related articles, similar past jobs, comparable matters.
- Feeding the right pages to a model. When a document library is too large for a context window, embeddings pick the sections worth including.
That last one is the important one commercially. It's how you get an AI assistant that answers from your 400-page operations manual without sending all 400 pages.
An Australian small-business example
A Perth manufacturer has fifteen years of job files. A new estimator wants to know whether they've quoted anything like a current enquiry before.
Keyword search fails — the old jobs use different product names, different abbreviations, and terminology that changed twice. Nothing matches what she types.
With the files embedded, a description of the current job surfaces four genuinely comparable past jobs, none of which share her wording. The knowledge was always there; it just wasn't reachable by anyone who didn't already know what it was called.
Where it goes wrong
- Related isn't the same as right. Embedding search returns what's similar. A superseded policy sits very close in meaning to the current one — which is why versioning matters more, not less.
- Chunking affects everything. Documents get split before embedding. Split badly and you get fragments missing the context that made them meaningful.
- It doesn't understand exceptions. A clause saying "this does not apply to contractors" is textually close to the rule it's excluding. Retrieval may surface one without the other.
All of which is a reason to keep a person reading the retrieved source rather than trusting a confident summary of it.
Why it's worth knowing the word
You'll never configure embeddings directly. But when someone proposes an AI system that "searches all our documents", knowing what's underneath tells you which questions to ask: how are documents split, how do you handle superseded versions, and does the answer show its sources?
Those three questions separate a document assistant that's genuinely useful from one that confidently quotes a policy you retired in 2021.
Frequently asked questions
Do I need to understand embeddings to use AI?
How is this different from a normal search?
Is this the same as RAG?
Where are the embeddings stored?
Does making embeddings send our documents somewhere?
Put this to work
Ad On Group runs AI training and enablement for Australian teams through Ad On AI — a three-month, self-paced program that takes non-technical staff from their first prompts to working AI agents.