Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.opper.ai/llms.txt

Use this file to discover all available pages before exploring further.

Control Plane features are in early access and need to be turned on per account. Contact support@opper.ai if you’re interested.
Guard protects your calls before content reaches the model and before responses reach your caller. You assemble Guard rules — each is either an LLM guard (a prompt-based classifier) or a Regex guard (pattern-based), with an action (Flag / Block / Redact) and a scope.

​
Rule kinds

Open platform.opper.ai and navigate to Controls → Guard to add a rule. There are two kinds.

​
LLM Guard

A custom prompt that classifies content. The model returns a flagged/not-flagged decision; what happens next depends on the action.
  • Prompt — free-text classification instructions, for example: “Flag support messages that include a plaintext password, API key, or access token. Do not flag placeholders or general security questions.”
  • Model — the model that runs the check:
    • Auto (default) — cheapest model that supports structured output.
    • Small — GPT OSS 120B, EU-hosted.
    • Medium — Kimi K2.5, EU-hosted.
    • Large — Claude Opus 4.6 — highest accuracy.
  • Action — Flag, Block, or Redact.
  • Input / Output / Input + Output — where the rule applies.
Quick-start templates fill the prompt and action with reasonable defaults for common checks: PII, Toxicity, Secrets. You can edit any of them after picking the template.

​
Regex Guard

Pattern-based detection. Use this for known formats you want to catch deterministically.
  • Patterns — a list, each with an optional name, the regex, and flags chosen from i, m, s, d, y, u. The editor includes a test harness that highlights matches against sample text, plus an “AI-generate” button that drafts a pattern from a description.
  • Replacement — text used by Redact (default ***).
  • Action — Flag, Block, or Redact.
  • Input / Output / Input + Output — where the rule applies.
Quick-start templates exist for Email, Phone, Credit Card, SSN, and API Key patterns.

​
Actions

ActionWhat happens
FlagThe match is logged on the trace; the call proceeds normally.
BlockThe call is rejected — before the model runs (Input) or before the response returns (Output).
RedactThe matched content is replaced — either with the default *** or with the replacement text you set.

​
Scope

  • Organization — every project and every function.
  • Projects — one or more specific projects.
  • Functions — specific functions inside a project.
You can layer rules — an org-wide secrets check plus stricter rules on customer-facing functions, for example. Each rule fires independently.

​
In traces

Guardrail events appear on a span with a ShieldCheck icon labeled Guardrail, with status:
  • Passed — the call satisfied the rule.
  • Flagged — the rule matched but the action was Flag, so the call proceeded.
  • Blocked — the rule rejected the call (red).
Each event carries the rule’s name (when set), a family label (“LLM guard” or “Regex guard”), and a scope badge (Org-level / Project-level / Function-level). Redacted content appears already replaced in the span’s input or output — there’s no separate visualization for the redaction.

​
In playground

Guard rules apply to playground calls when the Project controls toggle in the task sidebar is on. The guardrail event shows up on the trace — open the trace ↗ link from the output footer to inspect it. To run a call as if the guardrails weren’t in place, switch Project controls off for the run.
Start with Flag while you’re validating a new rule. Once you trust it, switch to Block or Redact for the same rule.
Combine an LLM Guard for fuzzy detection (e.g., toxicity) with a Regex Guard for exact patterns (e.g., credit card numbers). Regex is faster and cheaper; LLM Guard catches things regex can’t.