
Artificial intelligence · March 4, 2026 · Gynvix · 1 min read
What a retrieval system actually stores
A retrieval feature is a set of documents, chunks, and a rule for what the assistant is allowed to cite.
Documents before vectors
The useful object is the source document and the passage that was shown to a person. An embedding is an index. If you cannot open the passage, the answer is not reviewable.
Chunks are an editorial choice
Splitting a policy on every heading keeps a citation readable. Splitting on a fixed character count often cuts a rule in half. We write down the splitter.
Refusal is a feature
If no passage clears the threshold, the assistant should say it does not know. A fluent guess is a bug.
ts
function answer(passages: Passage[]) { if (passages.length === 0) return { refused: true, text: 'No approved source.' }; return { refused: false, text: draftFrom(passages), sources: passages };}