Sources index automatically when added. Most teams never need to manually re-train. But knowing how it works — and what the metrics mean — helps you spot quality issues early.
When does re-training happen automatically?
- When you add a new source
- When you edit a text or Q&A source
- When you re-crawl a URL source
- When you re-upload a file
Each happens in seconds. There's no batch nightly job — your changes are live as soon as the indicator turns green.
When to manually re-train
Click the Re-train button in the agent's Sources tab if:
- You've made many source changes at once and want to be sure
- You've changed the model and want a fresh index
- An indexing error appears in the source list
Knowledge metrics
Open the agent's Dashboard tab and look for the Knowledge section. You'll see:
- Sources count — total active sources
- Knowledge size — how many indexed passages (chunks)
- Last trained — when the index was last updated
- Coverage (when shown) — how often customer questions were grounded in your sources vs answered from the model's general knowledge
Tip
Coverage is the metric to watch. If it drops below 60%, your sources aren't keeping up with what customers ask. Time to add more.
Cleaning up
Over time you'll accumulate stale sources. A monthly review:
- Sort sources by date added (oldest first)
- For each: ask "is this still accurate?"
- Delete or replace anything that isn't
Source priority
When two sources contain conflicting information, the agent picks whichever is more relevant to the question — but you can't entirely predict which. The fix is to delete the wrong source so there's only one right answer in the index.