Finding things by position

If meaning is a position, then finding related things is finding nearby positions.

Think about how ordinary keyword search works: you type words, the system looks for documents containing those words. Type car repair costs and a document about automobile maintenance expenses may not come back at all, despite being exactly what you wanted. It shares no words with your question. To a system matching text, it is invisible.

Now do it with positions. The same trick that gives a word a position works on a whole run of text: feed in the sentence, or the document, and get back one position standing for the lot. So the question gets a position, every document gets a position, and which documents relate to this question turns into a question about distance: which positions sit nearest to this one?

The document that never used the word car still lands in the neighbourhood, because automobile and car were pulled together by keeping the same company, exactly as this chapter described.

Nothing new is needed for this. It is the same space, used for a different purpose. Meaning became geometry, and geometry can be searched.

The final module gives a model the ability to look things up rather than answer from memory alone. This is the machinery it uses.