Embedding

This is the machinery underneath every AI feature that finds things by meaning rather than by spelling. Searching your own documents and getting the right passage although it uses none of your words. An assistant answering from a knowledge base. A support tool spotting that two tickets describe the same fault. A shop suggesting a product like the one you looked at. All of them work this way. It also explains a class of failure worth recognising. Because closeness is judged on meaning as it appeared in training, a system can confidently retrieve something adjacent to what you meant, and adjacent is not the same as right.

A model reads the text and produces a fixed-length list of numbers for it, usually several hundred to a couple of thousand of them. The list is not readable and no single number in it corresponds to anything you could name. What matters is only its position relative to other lists produced by the same model, which is why these things are useful in comparison and meaningless in isolation.

The arrangement is learned rather than designed. During training the model is pushed to place text that appears in similar circumstances near to other text like it, so it ends up with an internal geography where related ideas cluster. Nobody chose which region holds complaints about delivery times; that region exists because the training material put those sentences in company with each other.

Comparison is arithmetic, which is what makes it fast. Given a question, the system produces its list, then measures how close that list is to the ones it has already stored and returns the nearest. Millions of stored items can be searched this way in a fraction of a second, and specialist stores exist to make it faster still.

Everything is done with the same model at both ends. A list produced by one model cannot be meaningfully compared with a list produced by another, because each has its own learned geography. This has a practical consequence people meet the hard way: changing the model means rebuilding every stored list, so it is not a setting you can flip.

Longer material gets divided before it is turned into lists. A whole report reduced to one position is too blunt to be useful, so text is split into passages first, and the size of those passages decides what the system can find. Passages that are too large return answers padded with irrelevance; passages that are too small lose the context that made them meaningful.

Closeness is not correctness. Two sentences can sit near each other because they share a subject while saying opposite things, which is why a search for a policy that permits something can return the passage forbidding it. The arithmetic has no notion of agreement or negation, only of proximity, and any system built on it needs a step that reads what came back.

Meaning-based retrieval and word matching fail in opposite directions, which is why serious systems run both. Proximity finds the passage that never uses your words and misses the exact product code you typed, because a code carries little meaning to sit near anything. Word matching does the reverse. Combining the two, then re-ranking what both return, covers considerably more ground than either alone, and is close to standard practice for that reason.

The same trick works beyond text. Images, audio and code can be placed in the same kind of arrangement, and where a model has been trained on more than one medium at once, a description and a photograph can end up near each other. That is how searching a picture library with a sentence works, and it is the same machinery rather than a separate feature.

• Ask a notebook tool a question using none of the words in your uploaded document and watch it surface the right passage anyway, which is proximity rather than word matching.

  NotebookLM

• Search a company workspace for a subject described differently by every team, and find the results agree with each other even though the wording does not.

  Glean

• Load a folder of documents into a self-hosted assistant, where the lists are produced and stored on your own machine and nothing is sent out to be compared.

  AnythingLLM

What do the numbers in an embedding actually mean?

Individually, nothing you could name. The arrangement is learned rather than designed, so no single position in the list corresponds to a readable property like formality or topic. Meaning lives in the relationships between whole lists, which is why they are only useful when compared with others from the same model.

Do I need to understand embeddings to use AI at work?

Not to use the features, no. It becomes worth understanding when you are choosing or troubleshooting a system that searches your own material, because the common failures, retrieved passages that are related but wrong, or a search that degrades after a model change, only make sense once you know what is being compared.

Are embeddings the same as how a model stores knowledge?

No. Embeddings are produced on demand from text you supply and stored by your system. What a model learned in training lives in its internal settings and is not something you can inspect, store or search. They are separate mechanisms that happen to both involve numbers.

Can someone reconstruct my documents from stored embeddings?

Not straightforwardly, since the process is not designed to be reversible. Research has shown that meaningful information can sometimes be recovered from stored lists, so treat them as derived from sensitive material rather than as anonymised. Where confidentiality matters, the storage deserves the same protection as the documents themselves.

Why did our document search get worse after an upgrade?

Most often because the model producing the lists changed and the stored ones were not rebuilt, leaving new questions compared against an older arrangement. The symptom is retrieval that looks almost right and is subtly off. Rebuilding the stored lists with the current model is the fix, and it is worth confirming before assuming the new model is worse.

What size should passages be before embedding them?

Large enough to stand alone and small enough to be about one thing. A paragraph or two is a common starting point. Too large and results arrive padded with irrelevance; too small and passages lose the context that made them meaningful. It is worth testing on your own material rather than adopting a figure.

• An embedding turns text or an image into a list of numbers whose position encodes meaning.
• Related material ends up nearby, which is what allows searching by meaning rather than by wording.
• The arrangement is learned in training, so lists from different models cannot be compared.
• Closeness is relatedness, not agreement: opposites often sit near each other.
• Longer material is split into passages first, and passage size decides what can be found.

• NotebookLM

  Answers come from uploaded documents, which is meaning-based retrieval in its simplest form.

• Glean

  Searches a whole workspace where every team words the same subject differently.

• AnythingLLM

  Self-hostable, so the lists can be produced and kept on your own machine.