Embeddings explained simply: how machines search by meaning
Embeddings, explained simply, turn text into a set of numbers where phrases with similar meanings get similar numbers. Search then compares numbers instead of words. A query for "time off without pay" leads straight to the policy clause on unpaid leave, even though the two share no words in common.
Why this matters
- It explains why search across company documents works by meaning rather than by matching words
- It clarifies what exactly an AI invoice is charging for, and where the cost can be cut
- It makes the difference clear between a vendor proposing something necessary and one proposing something superfluous
What is an embedding, in plain language
Picture a filing system where every document carries map coordinates instead of a title. Documents about leave land in one corner, procurement in another, workplace safety in a third. The computer never reads the text: it looks at which points sit close together.
Those coordinates are the embedding. A model reads a chunk of text and returns a list of numbers, typically several hundred to several thousand of them. Each number describes some facet of the meaning: the subject, the tone, the domain it belongs to. On their own the numbers tell a human nothing — only the distances between them matter.
From there it is all arithmetic. An employee asks how to take time off without pay, the question is converted into coordinates too, and the system finds the nearest points. Back comes the policy clause on unpaid leave, even though not a single word in the query matched it.
How it works
- Documents are cut into chunks. A hundred-page policy cannot be reduced to a single point — the meaning would smear. The text is split into fragments a few paragraphs long.
- Each chunk is run through the model. The output is a set of numbers for that fragment.
- The numbers go into a dedicated store — a vector database. Stored alongside them are the original text and labels: which document it came from, which revision, which department it applies to.
- The user's question is converted into the same kind of numbers by the same method.
- The nearest fragments are retrieved and handed to the model, which composes an answer in plain language.
One requirement is non-negotiable: the question and the documents must be converted into numbers by one and the same model. Different models build different maps, and coordinates from one map mean nothing on another.
An example from practice
At a company with distributed warehouses, storekeepers were hunting for answers in a three-hundred-page goods-receipt policy. Keyword search was no help: the document says "quantity discrepancy on receipt", while people ask what to do when less was delivered than ordered.
After the policy was converted into embeddings, the average time to find an answer fell from six minutes to twenty seconds. A side effect mattered more than the speed: the senior storekeeper stopped being called about every disputed case, which freed up roughly four hours of his week.
What the business gets out of it
- Search starts working the way people ask, not the way the document is written. That removes the main reason knowledge bases go unused.
- Answers rest on company documents, not on the model's general knowledge. This is the foundation of the arrangement known as RAG.
- It is cheap to run. The conversion into numbers happens once, when a document is loaded, not on every question. The expensive part is the model's answers, not the search.
When it can be skipped
If there are few documents and they are well structured, ordinary keyword search will do the job more cheaply and more predictably. The order of magnitude is this: with up to a few dozen documents under clear titles, there is no reason to build vector search.
Embeddings are no help either where exact matches are required: part numbers, contract numbers, error codes. Search by meaning is precisely the kind that substitutes something close — instead of part number 4517 it will find 4519. Such cases call for a hybrid: exact search on numbers plus semantic search on descriptions.
And separately: an embedding does not fix contradictions or superseded revisions — search answers from whatever sits in the store. That is why the list of documents currently in force is drawn up before loading, not after the first complaints: a few days of work that pay for themselves immediately.
What is worth checking
Quality depends on the model. One trained mainly on English draws finer distinctions less well in Russian text, especially in legal and technical wording. This has to be tested on the company's own documents and its own typical questions rather than on published rankings: every company has its own language, and the leader of a general benchmark may lose on a particular one.
Frequently asked questions
What is an embedding, in plain language?
It is a way of turning text into a set of numbers that preserves meaning. A computer does not understand words, but it is very good at comparing numbers. After that conversion, "unpaid leave" and "time off without pay" end up next to each other, even though they share no words at all.
How is an embedding different from a vector?
A vector is simply a list of numbers, a mathematical object. An embedding is a vector derived from text or an image that carries its meaning. Put another way, the vector is the notation and the embedding is what has been written in it.
What does it cost to turn documents into embeddings?
The price is based on the volume of text. The operation costs a fraction of what model answers cost: reprocessing the entire document base of a mid-sized company once is typically cheaper than a single month of running it in production. Exact figures come from the provider's per-million-token rates.
Do documents have to leave the company?
Not necessarily. Some embedding models are open and can be deployed inside a company's own perimeter. Documents then never leave it, and the bill covers a server rather than tokens.
How often do embeddings need to be recalculated?
Every time a document changes. If a policy is amended while the numbers still reflect the old version, search will confidently answer from outdated text — and no one will notice until someone checks by hand.
What is the difference between an embedding and a vector?
A vector is just a list of numbers — a way of writing things down. An embedding is a vector with a special property: the model picks its numbers so that texts with similar meaning get similar vectors. Every embedding is a vector, but not every vector is an embedding.
Let’s discuss your project?
Tell us about your process — we’ll suggest where AI pays off fastest.
Related articles
What Inference Is and What Makes Up the Bill for It
Inference is the moment a trained model answers a request — and it is what gets billed every month.
Contextual chunk enrichment — what it means in plain terms
Contextual chunk enrichment means adding a short explanatory note to every piece of a document so that knowledge base search finds the right one more often.