Grounding an assistant in docs that change weekly
On this page
The model choice for a retrieval-augmented assistant is rarely the bottleneck. The bottleneck is that internal docs change every week, and a stale index answers confidently with last month’s process.
Re-index on write, not on a schedule
Instead of a nightly re-index job, the assistant’s index updates when a doc is actually saved — a webhook from the wiki triggers a re-embed of just that page. The lag between “someone fixed the runbook” and “the assistant knows about it” dropped from up to 24 hours to a few seconds.
Always show the source
Every answer links back to the page it came from. That single decision did more for trust than any prompt tweak: support agents check the source before repeating the answer to a customer, which also means they catch a wrong or outdated doc immediately, instead of the assistant quietly repeating it forever.