In the previous post we talked about the gap between keyword search and semantic search, how one matches characters while the other compares meaning. But that raises an obvious question: how does a computer understand meaning at all?
The answer involves a small but powerful AI model, a list of numbers, and a map that doesn't exist in any physical space.
Text is hard for computers, numbers are easy
Computers are extraordinarily good at comparing numbers. They can tell you in microseconds whether two numbers are close or far apart, whether a sequence of numbers points in the same direction as another, and how similar two lists of numbers are to each other.
Text, on the other hand, is hard. "Brilliant" and "exceptional" look nothing alike as character sequences, yet a human immediately understands they mean nearly the same thing. "Bank" appears identical in "river bank" and "savings bank," yet means completely different things depending on context.
For decades, computer search avoided this problem by just not trying to solve it, hence keyword search. But a different approach became possible once AI models grew capable enough to learn the relationship between language and meaning.
The model that reads your documents
When you import a document into ThinkableSpace, a small AI model reads it. Not to summarise it, not to answer questions about it, just to understand what it's saying, at a deep enough level to represent that meaning as numbers.
The process works on chunks, short passages of a few hundred words at a time, rather than entire documents. This matters because a 50-page report contains many different ideas. Treating it as a single unit would blur all of them together. Breaking it into chunks lets the system represent each idea separately, so a search for one specific concept can find exactly the passage that discusses it, not just the document that happens to contain it somewhere.
Each chunk gets fed through the model. The model processes the language, considers how the words relate to each other and to the broader patterns it learned during training, and produces an output: a list of several hundred numbers.
This list of numbers is called an embedding. It's a compact, precise description of what that chunk of text means.
A map of meaning
Here's the key property that makes embeddings useful: the model was trained so that texts with similar meanings produce similar numbers.
Not similar words, similar meanings. "The meeting was cancelled" and "the call was called off" would produce embeddings that are close to each other. "The meeting was cancelled" and "a recipe for banana bread" would produce embeddings that are far apart.
You can think of it as a vast, invisible map. Every piece of text has a location on this map. Texts about similar topics cluster together. Texts about entirely different things are separated by distance.
This map has no physical dimensions you can visualise, it exists in mathematical space, but the underlying idea is intuitive: things that mean the same thing end up near each other.
How search uses this map
When you type a query, it goes through the same model. Your question, however loosely phrased, whatever words you happen to choose, gets converted into its own list of numbers and placed on the same map.
The search then finds the document chunks whose positions are closest to your query's position.
This is why vocabulary stops mattering. "Q4 budget discussion" and "quarterly financial review" land in the same part of the map. When you search for either phrase, you find documents near that region, regardless of which exact words those documents used.
The model acts as a translator. You speak in your words. Your documents were written in theirs. The embedding converts both into the same language, numbers, and from there, finding relevant results becomes a geometry problem.
This happens entirely on your device
The model that does all of this runs locally. It was downloaded to your machine when you first set up ThinkableSpace, and it runs on your CPU or GPU whenever you import documents or run a search.
Your text is never sent anywhere to be understood. The intelligence lives on your device.
This is one of the reasons the model was chosen carefully for size: small enough to download once and run quickly on ordinary hardware, capable enough to genuinely understand language rather than just approximate it.
Every document you've ever imported has been read, understood, and translated into this numerical representation, stored quietly on your drive, ready to be searched in milliseconds.
Next: how ThinkableSpace searches through millions of these numerical descriptions in under 30 milliseconds, without checking each one individually.