Documentation as Findability in AI-Assisted Systems

In embedding-based retrieval, what you don’t describe, AI can’t find.

This has always been true for human search, poor metadata means poor discoverability. But AI-assisted systems amplify the effect. When semantic similarity is the retrieval mechanism, the text is the thing. Undocumented concepts don’t just rank lower; they become invisible.

The implication for knowledge engineering: documentation isn’t a nice-to-have layer on top of structure. It’s the interface through which AI systems access that structure.

Well-designed schemas with sparse annotations perform worse in similarity-based retrieval than mediocre schemas with rich descriptions. Quality of structure and quality of documentation are now entangled in ways they weren’t before.

Related: 06-atom—annotation-quality-bias