Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt

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

Use this recipe when you need to protect one or more API keys with budget limits, routing restrictions, prompt-injection checks, or sensitive-info handling.

​
1. Create the guardrail under the workspace

Create guardrails at the workspace level first so the same policy can be reused across multiple keys. Recommended first pass:
  • add model and provider restrictions before content checks
  • keep the first rollout narrow to one key or one environment
  • name the guardrail after the policy outcome, not the team name

​
2. Attach it to a target API key

From the API key settings page, attach the guardrail to the key that should receive enforcement. Use this sequence:
  1. start with a staging or low-risk key
  2. attach one guardrail at a time when possible
  3. confirm the key detail dialog shows the applied guardrail

​
3. Set budget policy and routing restrictions

Common combination:
  • daily cost limit
  • provider blocklist for low-trust providers
  • model allowlist for the exact production models that should remain routable
Use the live preview in the guardrail editor to verify:
  • which providers stay allowed
  • which models become blocked
  • whether Only allow is narrower than intended

​
4. Add prompt-injection and sensitive-info rules

For a practical rollout:
  • start prompt injection with flag or redact
  • move to block only after reviewing false positives
  • use deterministic sensitive-info rules first:
    • email
    • phone
    • SSN
    • credit card
    • IP address
If you need custom patterns, add them as explicit regex rules and test them in the preview input before saving.

​
5. Test the policy before broad rollout

Use the guardrail preview with:
  • one benign input
  • one clearly disallowed input
  • one realistic production-like prompt
Look for:
  • correct redaction text
  • correct blocked/allowed behavior
  • no accidental matches on unrelated content

​
6. Verify enforcement in activity and logs

After sending real requests through the guarded key, verify:
  • the Guardrail Enforcement panel on the activity page
  • request detail dialogs for blocked or redacted requests
  • the API key detail dialog for per-key guardrail activity
You should see:
  • blocked counts
  • redacted counts
  • flagged counts
  • the specific guardrail and detector details when enforcement happened

​
7. Expand rollout safely

When the first key behaves correctly:
  1. attach the same guardrail to more keys
  2. tighten actions from flag to redact or block
  3. add higher-latency name and address detection only where the latency tradeoff is acceptable
Last modified on May 19, 2026