Skip to content

docs: write safety guardrails how-to for Guardian Intrinsics API #802

Description

@planetf1

Context

The mellea-website homepage features a "Safety & Guardrails" card linking to /how-to/safety-guardrails. That page doesn't exist — #788 added a temporary redirect to /advanced/security-and-taint-tracking.

However, the existing advanced page only documents the deprecated GuardianCheck requirement API (deprecated in v0.4). The current Guardian Intrinsics API in mellea/stdlib/components/intrinsic/guardian.py has zero documentation.

What needs to happen

  1. Write /how-to/safety-guardrails covering the current Intrinsics API:

    • guardian_check() — harm, jailbreak, profanity, etc.
    • policy_guardrails() — custom policy criteria
    • factuality_detection() / factuality_correction() — groundedness
    • CRITERIA_BANK pre-baked criteria (10 entries)
    • Plugin-layer guardrails pattern (PII redaction via plugins)
  2. Update or retire advanced/security-and-taint-tracking.md:

    • Currently misnamed ("taint tracking" has no content)
    • Documents only the deprecated GuardianCheck API
    • Either add a deprecation banner + redirect to the new page, or fold remaining useful content into it
  3. Remove the redirect in docs.json once the real page exists

  4. Fix pip install mellea reference in the advanced page → uv

References

  • https://docs.mellea.ai/how-to/safety-guardrails is a 404 #788 — original 404 report
  • mellea/stdlib/components/intrinsic/guardian.py — current API source
  • mellea/stdlib/requirements/safety/guardian.py — deprecated API
  • docs/docs/tutorials/04-making-agents-reliable.md — existing Guardian tutorial content (steps 4-7)

Metadata

Metadata

Assignees

Labels

area/adapter-functionsGranite adapter functions: framework and adaptiers including RAG, Guardian, Corearea/stdlibCore abstractions: Context, MOT, SamplingStrategy, formatters, serializationdocumentationImprovements or additions to documentation

Type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions