Skip to content

docs: rewrite guardrails for the Bedrock-backed feature#63

Merged
micheleRP merged 3 commits into
mainfrom
michele/guardrails-bedrock-rewrite
Jun 5, 2026
Merged

docs: rewrite guardrails for the Bedrock-backed feature#63
micheleRP merged 3 commits into
mainfrom
michele/guardrails-bedrock-rewrite

Conversation

@micheleRP
Copy link
Copy Markdown
Contributor

@micheleRP micheleRP commented Jun 5, 2026

What

The guardrails docs described a feature that doesn't ship: PII / Toxicity / Custom-webhook evaluators, a global INPUT/OUTPUT/BOTH phase selector, attach-in-create, and a violations UI. This rewrites all four pages to match what actually ships, verified against guardrail.proto on cloudv2 origin/main plus the ADP /guardrails UI.

What ships (and is now documented)

  • AWS Bedrock Guardrails-backed. Every guardrail needs an AWS region + Bedrock credentials, even to protect non-Bedrock providers (wrapped via Bedrock's ApplyGuardrail).
  • 3-step create wizard: Details → Backend → Policies, with a per-policy "Evaluate this policy" staging toggle.
  • Six optional policies: Content filters, Word filters, Denied topics, Sensitive information (PII), Contextual grounding (output-only), Automated reasoning (detect-only).
  • Provider-side binding (a provider references one guardrail; reusable across providers).
  • No violations UI: blocked requests return the configured message; activity is recorded on the OpenTelemetry trace and read in transcripts. Streaming output evaluation is skipped.

Pages

  • guardrails/overview.adoc — reframed around the six policies, Bedrock requirement, evaluation model, binding, and blocking behavior.
  • guardrails/create-guardrail.adoc — rewritten to the 3-step wizard with real fields.
  • guardrails/types-reference.adoc — retitled Guardrail Policy Reference; full per-policy schemas. Nav + index updated.
  • guardrails/violations.adoc — reframed/trimmed to how a blocked request surfaces.
  • budgets.adoc — corrected the guardrail-cost section (guardrail usage is billed by AWS Bedrock, not tracked in ADP spend) and fixed stale link text.

Preview

🤖 Generated with Claude Code

The guardrails pages described a PII/Toxicity/Custom-webhook model with a
violations UI that never shipped. Rewrite all four pages to match what
ships, verified against guardrail.proto on cloudv2 origin/main and the
ADP UI:

- overview: reframe around the six Bedrock policies (content filters,
  word filters, denied topics, sensitive information, contextual
  grounding, automated reasoning), the AWS Bedrock requirement, input/
  output evaluation with the streaming-output caveat, provider binding,
  and how blocked requests surface.
- create-guardrail: rewrite to the three-step wizard (Details, Backend,
  Policies) with the per-policy 'Evaluate this policy' staging toggle,
  credential source/type, provider attachment, and edit/disable/delete.
- types-reference -> 'Guardrail Policy Reference': full per-policy schemas
  (categories, strengths, actions, PII entity types, regex, topics,
  grounding thresholds, automated-reasoning versioned ARNs).
- violations: reframe to how a blocked request surfaces (blocked message,
  anonymize, OTel trace/transcript); drop the violations-table, metrics,
  and fail-open/closed claims that do not ship.
- budgets: correct the guardrail-cost section (guardrail usage is billed
  by AWS Bedrock, not tracked in ADP spend) and fix stale link text.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@micheleRP micheleRP requested a review from a team as a code owner June 5, 2026 19:58
@netlify
Copy link
Copy Markdown

netlify Bot commented Jun 5, 2026

Deploy Preview for redpanda-agentic-data-plane ready!

Name Link
🔨 Latest commit d8ba282
🔍 Latest deploy log https://app.netlify.com/projects/redpanda-agentic-data-plane/deploys/6a232fae2b2c230008d9d0d1
😎 Deploy Preview https://deploy-preview-63--redpanda-agentic-data-plane.netlify.app
📱 Preview on mobile
Toggle QR Code...

QR Code

Use your smartphone camera to open QR code link.

To edit notification comments on pull requests, go to your Netlify project configuration.

micheleRP and others added 2 commits June 5, 2026 14:07
The roadmap TODO listed configurable caps and per-agent caps as not yet
shipped, which contradicts the page body (both ship). Align it with the
NOTE on the page: only shared pools, model/provider/org-scoped limits,
and notification delivery remain unshipped.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Remove step numbers from create-guardrail headings (style rule).
- Spell out RAG and ARN on first use; drop 'should'.
- Standardize the violations learning-objective lead-in.
- Link the transcript glossterm on first mention.
- Prefer active voice in guardrail/budget descriptions.
- Trim the overview description under 155 characters and rename the
  policy-types heading.
- Drop the hardcoded policy count ('six') so it can't go stale.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@micheleRP micheleRP merged commit 427075f into main Jun 5, 2026
4 checks passed
@micheleRP micheleRP deleted the michele/guardrails-bedrock-rewrite branch June 5, 2026 20:25
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant