Guardrails
Layered policies for PII, prompt injection, topics, toxicity, and bias.
Guardrails screen every inference request (/v1/chat/completions, /v1/messages, /v1/responses) after routing and before any provider sees it. Policies are managed from the Guardrails page.
Only user-authored text is scanned: the text parts of user messages. System prompts, assistant turns, tool results, and model output are not.
Layered, most-specific-wins
Policies can be attached at five scopes — global, provider, model, chain, and key. A non-global policy applies only when its target matches the request:
| Scope | Target matches |
|---|---|
key | the API key's id |
chain | the chain name the request resolved through |
model | the requested model, or any candidate target's model / provider/model / alias |
provider | any candidate target's provider slug |
For a single request, the router resolves the effective setting per detector with a most-specific-wins rule: key beats chain, which beats model, then provider, then global. The most specific policy that enables a detector supplies its settings.
The consequence: a global guardrail cannot be turned off from a key — keys can only tighten or tune, never loosen the layers above them. Provider and model policies apply when they match any fallback target, so fallback never escapes them.
Detectors
| Detector | What it checks |
|---|---|
pii | Emails, IPs, phone numbers, credit cards, IBANs, and Indonesian identities (NIK, NPWP, passport) — with masking strategies |
injection | Prompt injection patterns, tiered |
topics | Allowed or blocked topic lists (allow/block mode) |
toxicity | Rude/harmful content with score thresholds |
bias | Bias with score thresholds |
Each detector has enabled and action:
block— reject the request with400before dispatch.warn— forward it and flag the response.log— forward it and only record it.
PII has a masking strategy instead (redact → <PII>, mask → ****, hash → [hash:…]). Matches are rewritten in the user text before it is forwarded, even when another detector's warn outranks it.
Any decision other than allow adds an X-TeraRouter-Guardrails: <decision> response header and a guardrail.<decision> entry to the Audit Logs tab. A failed policy lookup fails closed with 500.
Testing without risk
The policy editor has a Test Policy panel (POST /v1/guardrails/evaluate): send a sample text, see which detectors trigger, the decision, and the masked text — without sending anything to a provider. Omit config to test against every enabled policy merged together.
curl -X POST http://localhost:8080/v1/guardrails/evaluate \
-H "Authorization: Bearer <dashboard-jwt>" \
-H "Content-Type: application/json" \
-d '{ "text": "my email is budi@example.com" }'Import and export
Export all downloads guardrails-policies.json ({ "policies": [...] }). Import validates the whole file first, then creates every policy in it alongside the existing ones.
Detection credibility
The built-in detectors are deterministic (regex/lexicons) and run inside the router process. No external engine (Presidio, OpenAI moderation) is wired in, so the external_detectors toggle currently only changes the engine notes shown by the test panel.