Main idea

Searching for "regulatory maintenance" in a vector database and getting back the sheets on periodic inspections, certifications and the NF X 60-010 standard: none of them contains the phrase you looked for. The match was made on neighbourhood of meaning, not on coincidence of words.

What this removes from the person searching is precise: the obligation to know the exact vocabulary under which the information was written. A keyword engine requires the question to use the terms of the answer — hence that you already know how the answer was phrased. That condition falls away.

The property holds for code as well: indexing a whole repository this way finds a rule buried in the middle of a file, with nothing in its name to flag it, that no keyword search would have brought back. That is the tool's best use — and its limit rests on the same mechanism, since an engine that ranks by proximity always answers and never says there is nothing.

Layer added by "I wrote a product's ontology. Three times, I thought I was done." (2026-08-11).

Why it matters

This changes what you demand of the way notes are written. The title no longer has to contain every term someone might one day search under, and the time spent guessing those terms can go elsewhere.

Above all it makes usable a base you did not write yourself, or have forgotten: you can search it without knowing its plan, which is the condition for it to outlive its author.

Nuances and limits

Proximity of meaning is not relevance. Two notes that are neighbours in word space may answer different questions, and the final sorting is still left to the reader.

And the procedure fails exactly where exact coincidence was the right criterion: telling two close standards apart, retrieving a numerical threshold, isolating a date. Semantic neighbourhood erases the differences of detail, which are sometimes the whole point.

Open questions

• How do you notice that a semantic search has silently discarded the one note that answered?