Guardrails API

Endpoint for evaluating content against guardrail rules.

Evaluate

POST /api/client/v1/guardrails/evaluate

Request

{
  "guardrail_key": "pii-checker",
  "text": "My email is john@example.com and my phone is 555-0100",
  "target": "input"
}

Parameters

Field	Type	Required	Description
guardrail_key	string	Yes	Key of the guardrail to evaluate
text	string	Yes	Content to evaluate
target	string	No	input, output, or both (default: input)

Response

{
  "passed": false,
  "guardrail_key": "pii-checker",
  "guardrail_name": "PII Checker",
  "action": "flag",
  "findings": [
    { "category": "email", "message": "Email address detected", "block": false },
    { "category": "phone", "message": "Phone number detected", "block": false }
  ],
  "message": null
}

Response Fields

Field	Description
passed	true if no findings triggered, false otherwise
guardrail_key	Key of the evaluated guardrail
guardrail_name	Display name
action	Configured action: block, flag, or redact
findings	Array of detected issues
message	Optional message for blocked content

Guardrail Types

Type	Evaluation Method
PII Detection	Regex-based pattern matching (15 categories)
Content Moderation	Category-based content evaluation
Prompt Shield	Prompt injection detection
Custom Prompt	LLM-based evaluation with custom rules

Inference Integration

Guardrails can be attached to models and evaluated automatically during chat completions. When a guardrail blocks a request, the chat API returns:

{
  "error": {
    "type": "guardrail_block",
    "guardrail_key": "pii-checker",
    "action": "block",
    "findings": [...]
  }
}

Errors

Status	Description
400	Missing guardrail_key or text
401	Invalid API token
404	Guardrail not found