Unify Logo Footer.svg
Enterprise Search
Logo
Controls

Controls

Logo

4 mins read READ

Controls are where you define and manage the guardrails that protect Enterprise Search, across Knowledge and Chat.

What is a control

A control is one reusable guardrail: it defines what it reads, how it checks, and what happens when it catches something. Controls live under Governance › Controls, with separate Chat and Knowledge tabs. The list shows each control's Name, Check, Applied on, Action, Owner, and Last updated, plus a toggle to turn it on or off inline. Use New Control to build one.

The control editor

The editor walks you through three parts:

  • Details — a name and description.

  • Check — what to look for, either a prebuilt check or a custom one you write yourself.

  • Action — what to do on a hit: Block or Flag on Chat, Restrict or Flag on Knowledge. Some Chat checks also offer Mask.

You author Chat and Knowledge controls on their own tabs — the surface is set by where you create it, and can't be changed afterward.

Checks and inputs

You either pick a prebuilt check or write a custom (LLM) check. A control's Inputs decide what it looks at — on Chat that's user input, agent output, tool name, tool input, or tool output; on Knowledge it's the indexed knowledge content itself. Prebuilt check options differ by surface, since Chat and Knowledge guard different things.

Custom (LLM) checks

Choose the Custom check to have a model judge the content in your own words. You give it the Inputs to read, a Check model that evaluates, a Check prompt describing what counts as a violation, and an Output shape — Binary (Yes/No) for a plain breach determination, or Scale (0–100) with a threshold, so you only act above a confidence you set. Use a custom check when no prebuilt one fits, such as an organization-specific policy phrased in natural language.

Severity and testing

Check severity is off by default; turn it on to label each hit's seriousness, either graded by the model or a fixed level you set. Severity then shows on the resulting incidents so you can prioritize. The Test this control panel lets you enter sample inputs and Run test to see whether the control would fire. New controls start Enabled, so switch Enabled off in the editor header if you want to test before the control takes effect — that way you don't block or restrict real traffic by mistake. Every control has an Enabled toggle, in the editor header and inline on the list.

What's in This Section

TabWhat it covers
ChatControls that check live conversations and tool use, and can Block, Flag or, for checks that support it, Mask on a hit.
KnowledgeControls that check indexed content, and can Restrict or Flag on a hit.