Staleness is the quiet failure
Nothing errors when an assistant answers from a superseded policy. It is confident, well-written, cited, and wrong — and it stays wrong until a customer or a colleague notices.
That is why currency needs engineering attention rather than good intentions.
Sync, do not rebuild
- Watch the document source for changes, rather than reindexing everything nightly
- Update the affected passages within minutes of a change
- Remove deleted documents from the index immediately
- Log every index change, so a wrong answer can be traced to a document version
A nightly full rebuild sounds simpler and it means your assistant can be a day out of date on the morning a price changes.
Down-weight the old
Where you cannot remove something, use the review date. Content not reviewed in two years should rank below current material and should carry a visible note in the answer.
It is a small piece of logic and it prevents a whole category of embarrassment.
The ownership model
- Every core document has a named owner
- Every core document has a review date, enforced by a reminder
- Changes to price, policy or process trigger a documentation check
- The unanswered-question log is reviewed monthly and turned into content
Prove it periodically
A short set of questions with known current answers, run weekly against the live system. If any answer changes unexpectedly, something upstream has moved.
It takes an hour to build and it is the difference between finding a staleness problem yourself and having a customer find it.