What an embedding is, and why two sentences with no shared words land together

"Can I get my money back?" and "Refunds are available within 30 days" have no word in common except "can" and nothing useful to match on. A customer sees the same question twice. A keyword index sees two unrelated strings.

An embedding is what closes that gap, and it is less mysterious than it sounds. It turns a piece of text into a position, and once everything is a position, "related" becomes a measurable distance.

The map, and where a question lands

Pick a question. The map stays still; only the question moves.

The map above has two dimensions because a page does. A real embedding has many more, which is the only part you have to take on trust. Everything else, including the fact that the Spanish question lands beside the English passages, works exactly as it looks.

What the numbers actually are

An embedding model reads a passage and produces a long list of numbers. Not words, not a summary, not a hash: coordinates.

The model was trained on enough text that passages used in similar ways end up with similar coordinates. Nobody wrote a rule saying "money back" relates to "refund". It fell out of the training, which is also why the thing occasionally has opinions you did not ask for.

Three properties follow, and they explain most of what you will see day to day.

• The same text always gives the same numbers. Embedding is deterministic, so re-indexing unchanged content changes nothing.
• Numbers from one model mean nothing to another. The coordinates are relative to that model's space, so changing models means re-embedding everything.
• Length does not change the format. A sentence and a page both become a list of the same size, which is exactly why a page averaged into one point is a worse point than four focused passages.

Why distance is not relevance

The number next to a passage is a similarity, and the honest way to read it is as a ranking signal rather than a verdict.

Two things trip people up. The first is that similarity always returns something: ask about the weather and you get your least unrelated passage, which is why a threshold exists and why an assistant needs permission to say it found nothing.

The second is scale. In this product, genuine matches land roughly between 0.33 and 0.58. A score of 0.45 is a good match, not a mediocre one, and anybody who sets a threshold at 0.8 because it sounds rigorous has just switched their retrieval off.

The free gift: languages

The second question in the map is the one worth dwelling on. "¿Puedo devolverlo?" lands beside English passages about refunds because a multilingual embedding model puts meaning in the same place regardless of the language carrying it.

For anybody selling in more than one country, that is the difference between maintaining one knowledge base and maintaining four. Write the refund policy once in English, and a question in Spanish or Portuguese still finds it; the answer comes back in the language the visitor used.

It is not magic and it is not uniform: the effect is strongest in the languages a model saw most, and a term of art or a product name can behave oddly in any of them. But the default is far better than most teams expect before they try it.

What this changes about your content

If retrieval is geometry, then writing is placing things on a map.

A page about four subjects is one point in the middle of four neighbourhoods, close to nothing. Four short pages are four points, each near the questions it answers. That is the mechanical reason behind the advice to keep one topic per page, and it is why that advice does more for answer quality than any setting on the retrieval tab.

Questions people ask

The practical takeaway is small and unglamorous. Your content is a set of points, your customers' questions are points too, and the work is making sure that somebody who asks in their own words lands next to the paragraph that answers them.