A Knowledge control is a guardrail that scans your indexed content before it becomes searchable, and restricts or flags on a hit.
What a Knowledge control reads
A Knowledge control's only input is Knowledge content — the indexed material itself. Unlike Chat controls, there's no separate input to choose; the check always runs against the content as it's indexed. For example, a PII check on knowledge content keeps personal data out of the index in the first place, rather than catching it after the fact in a reply.
Prebuilt checks
| Check | What it catches |
| PII detection | Personally identifiable information in documents. |
| Credentials & Secrets | API keys, passwords, and similar secrets left in indexed material. |
| Prompt Injection | Content crafted to hijack the agent when it's later retrieved. |
| Regex pattern | Custom pattern matches. |
| Custom terms | Specific terms you define. |
Knowledge offers a smaller set of prebuilt checks than Chat — there's no Content Filter, Denied Topics, or Blocked Tools option, since those are conversational concepts. Write a Custom check instead when you need a model to judge document content against a prompt you write, with a Binary or Scale output.
What happens on a hit
A Knowledge control's Action decides the outcome: Restrict hides the matching content from search results, and Flag allows it to stay searchable but logs an incident for review. Every action, hard or soft, creates an incident you can review under Governance › Incidents › Knowledge.
Severity and testing
Turn on Check severity to label hits as they're found, either graded by the model or at a fixed level, so incidents can be prioritized. Before enabling a new Knowledge control, use Test this control to enter sample content and confirm it fires the way you expect — this avoids restricting legitimate content from search results by mistake.