Skip to content

Safety Guardrails

The Safety Guardrails engine (tuluat-guardrails) secures both REST API calls and DAG workflow steps with automated input sanitization and output validation.

1. Guardrail Pipeline (GuardrailPipeline)

graph LR
    Input[Prompt Input] --> PII[PII Masking Filter]
    PII --> Injection[Prompt Injection Filter]
    Injection --> Inference[LLM Model Inference]
    Inference --> Validation[Output Validation Filter]
    Validation --> Output[Sanitized & Validated Output]

2. Guardrail Filters

2.1 PII Masking Filter (PiiMaskingFilter)

  • Automatically detects and replaces sensitive personal data in user prompts before calling LLMs.
  • Modes: EMAIL, CREDIT_CARD, SSN, PHONE.
  • Replacement: Replaced with customizable tokens (e.g. [REDACTED]).

2.2 Prompt Injection Filter (PromptInjectionFilter)

  • Inspects incoming prompts for adversarial jailbreaks, system instruction overrides, and prompt injection attempts.
  • Strategies:
  • BLOCK: Blocks execution and returns AgentResponse.blocked().
  • SANITIZE: Strips suspicious instruction phrases.

2.3 Output Validation Filter (OutputValidationFilter)

  • Validates model output against node-declared JSON Schema specifications.
  • Assesses confidence score (e.g. minConfidence: 0.7) and rejects malformed or unparseable JSON outputs.