Guardrails Service

slm-guardrails is the external content-safety service used by zammad-ai-workflow.

Purpose

It evaluates user prompts and generated responses before the workflow continues.

Endpoints

• GET /healthz for liveness
• GET /ready for model readiness
• GET /metrics for Prometheus metrics
• POST /api/v1/guardrails/prompt
• POST /api/v1/guardrails/response
• GET /api/v1/models

Request Flow

• Requests are optionally protected by a bearer token.
• Empty text is treated as safe.
• The service loads guardrail models on startup and fails fast if none can be loaded.
• The workflow client sends the configured model id and confidence threshold.

Configuration

The service uses SLM_GUARDRAIL_ as the environment prefix.

Relevant settings:

• api.host
• api.port
• api.auth_token
• api.max_input_length
• guardrails.confidence_threshold
• guardrails.huggingface_cache_dir
• guardrails.offline_mode
• guardrails.max_concurrency
• guardrails.default_model
• guardrails.models

Response Shape

Prompt evaluation returns:

• prompt_safety
• prompt_toxicity
• jailbreak_detection

Response evaluation returns:

• response_safety
• response_toxicity
• response_refusal

Integration

The workflow service uses app/guardrails/http_client.py to call this service when guardrails.enabled is true.