Guardrails API

Endpoint for evaluating content against guardrail rules.

Hook plane (September 2026)

Guardrails now attach at six lifecycle hooks (prompt.pre, input.pre, output.pre, output.stream.delta, tool.pre, tool.post) across nine policy families, each with its own enforce/monitor mode. The /evaluate contract below is unchanged; see the Guardrails guide for the hook model and the full family list.

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"
}

Parameters

Field	Type	Required	Description
guardrail_key	string	Yes	Key of the guardrail to evaluate
text	string	Yes	Content to evaluate

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

Policy Families

Family	Evaluation Method
Personal data (PII)	Regex-based pattern matching (15 categories)
Credentials	Pattern + entropy matching
Word Filter	Deterministic list/regex matching
Moderation	LLM classifier
Prompt Shield	LLM classifier (prompt injection detection)
Tool Access	Deterministic tool allow/deny policy

A guardrail's type (preset vs custom) is separate from its policy families — custom runs an LLM judged against your own rule text instead of the bundled families above.

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