Guardrail Concepts

Understand probes, profiles, rules, and deployment behavior

Guardrail Request Lifecycle

Guardrails can evaluate content at multiple stages before and after model processing.

Core Building Blocks

Providers

Providers supply detection capabilities. Bud AI Foundry supports Bud Sentinel and cloud moderation providers.

Probes

A probe targets a specific risk area (for example, PII, toxicity, jailbreak, or safety categories).

Rules

Rules are the detection units within probes. You can often enable/disable or tune behavior per rule.

Profiles

A profile is your reusable policy package:

• selected probes
• guard type scope
• severity threshold and overrides
• deployment targets

Deployments

Deployment activates the profile in runtime paths, either as standalone or attached to an existing endpoint.

Guard Types and Evaluation Scope

Guard Type	Evaluation Point	Common Intent
input	Before prompt reaches model	Block malicious input
output	Before response reaches user	Prevent unsafe output
retrieval/tools (where applicable)	During extended workflow stages	Govern tool/data paths

Severity and Policy Tuning

Higher thresholds generally increase sensitivity and can reduce unsafe leakage, but may increase false positives. Calibrate with real prompts and staged rollout.

Execution Patterns

Pattern	Behavior	Use Case
Parallel probe execution	Evaluate multiple probes concurrently	Lower latency under broad coverage
Sequential evaluation	Process probes in a fixed order	Controlled policy precedence
Fail-fast handling	Stop at first critical violation	Strict compliance gates
Aggregate handling	Collect all violations	Rich audit and analysis

Operational Views in Bud Admin

• Guardrails page: profile discovery, filtering, and creation entry point.
• Guardrail details: General, Probes, and Deployments tabs.
• Project Guardrails tab: project-scoped list with status and management actions.