skip to content
exer//notes · ai

the retrieval step everyone skips, and the week it costs later

29 May 2026

Chunking is not the hard part. Deciding what a document is turns out to be the hard part.

[ placeholder content ] — shipped with the site so the editorial layout can be reviewed. Not a published note.

Retrieval tutorials start at chunking. Real retrieval starts about four steps earlier, at a question nobody enjoys: what counts as a document?

A 400-page policy PDF is not a document. It is roughly ninety documents that happen to share a cover page, and treating it as one is why your top result is a table of contents.

The week it costs

You find out at the demo. Someone asks a specific question, the system returns something adjacent and confident, and the room quietly stops believing the rest of what you built.

The fix is not a better embedding model. It is a boring afternoon spent deciding what a retrievable unit is for this corpus, and a second boring afternoon writing the splitter that produces it.

Placeholder note shipped with the site so the editorial layout can be reviewed with realistic text. Replace before launch.