What it does

Knowledge sources are the data an agent grounds its generative answers on. You attach SharePoint sites, Dataverse tables, public websites, uploaded documents, and other connectors, and the agent reasons over them.

Key facts

• SharePoint and OneDrive grounding respects the signed-in user's existing permissions.
• Public website grounding is scoped by URL; it reads public pages, not authenticated ones.
• Dataverse lets the agent answer over structured business data, not just documents.
• Each source type has size, count and format limits — check current caps rather than assuming.

When to use / skip

Attach knowledge whenever you want the agent grounded in your content instead of the model's general training. Skip a source if its content is stale, contradictory or ungoverned — it drags answer quality down more than it helps.

Configuration decisions

• Which sources to attach, and their trust and priority order.
• Whether public web grounding is permitted at all.
• How and when source content gets refreshed and reindexed.

Gotchas

• Oversharing surfaces instantly. If SharePoint permissions are loose, the agent makes that obvious.
• More sources isn't better — noisy or overlapping sources confuse retrieval and dilute answers.

Consultant notes

• Source curation and permissions review is the real work here; the wiring is trivial by comparison.
• Prefer a small set of authoritative sources over a broad dump of everything.
• Coordinate with the M365 Copilot oversharing remediation — same underlying problem, same fix.

---

Recheck supported source types and limits each release.