Built-in Guardrails Overview

Pydantic AI Guardrails includes 16 ready-to-use guardrails for common validation scenarios. All guardrails are factory functions that return configured InputGuardrail or OutputGuardrail instances.

Input Guardrails

Input guardrails validate user prompts before they reach the LLM.

Guardrail	Purpose	Import
`length_limit()`	Limit prompt length by chars or tokens	`guardrails.input`
`pii_detector()`	Detect PII (email, phone, SSN, etc.)	`guardrails.input`
`prompt_injection()`	Detect prompt injection attacks	`guardrails.input`
`toxicity_detector()`	Detect toxic/harmful content	`guardrails.input`
`blocked_keywords()`	Block specific words or phrases	`guardrails.input`
`rate_limiter()`	Rate limit requests	`guardrails.input`

Quick Import

from pydantic_ai_guardrails.guardrails.input import (
    length_limit,
    pii_detector,
    prompt_injection,
    toxicity_detector,
    blocked_keywords,
    rate_limiter,
)

Output Guardrails

Output guardrails validate LLM responses after generation but before returning to users.

Guardrail	Purpose	Import
`secret_redaction()`	Detect leaked secrets/API keys	`guardrails.output`
`llm_judge()`	LLM-as-a-judge quality evaluation	`guardrails.output`
`json_validator()`	Validate JSON output structure	`guardrails.output`
`regex_match()`	Match output against regex patterns	`guardrails.output`
`no_refusals()`	Detect when LLM refuses to answer	`guardrails.output`
`min_length()`	Ensure minimum response length	`guardrails.output`
`require_tool_use()`	Ensure specific tools were called	`guardrails.output`
`tool_allowlist()`	Restrict which tools can be called	`guardrails.output`
`validate_tool_parameters()`	Validate tool call arguments	`guardrails.output`
`hallucination_detector()`	Detect potential hallucinations	`guardrails.output`

Quick Import

from pydantic_ai_guardrails.guardrails.output import (
    secret_redaction,
    llm_judge,
    json_validator,
    regex_match,
    no_refusals,
    min_length,
    require_tool_use,
    tool_allowlist,
    validate_tool_parameters,
    hallucination_detector,
)

Usage Example

from pydantic_ai import Agent
from pydantic_ai_guardrails import GuardedAgent
from pydantic_ai_guardrails.guardrails.input import (
    length_limit,
    pii_detector,
    prompt_injection,
)
from pydantic_ai_guardrails.guardrails.output import (
    secret_redaction,
    llm_judge,
)

agent = Agent('openai:gpt-4o')

guarded_agent = GuardedAgent(
    agent,
    input_guardrails=[
        length_limit(max_chars=2000),
        pii_detector(),
        prompt_injection(sensitivity='high'),
    ],
    output_guardrails=[
        secret_redaction(),
        llm_judge(criteria='Is the response helpful?', threshold=0.7),
    ],
)

Choosing the Right Guardrails

For Security

Concern	Recommended Guardrails
Prompt injection	`prompt_injection()`, `blocked_keywords()`
Data leakage	`pii_detector()`, `secret_redaction()`
Abuse prevention	`rate_limiter()`, `length_limit()`, `toxicity_detector()`
Tool safety	`tool_allowlist()`, `validate_tool_parameters()`

For Quality

Concern	Recommended Guardrails
Response quality	`llm_judge()`, `min_length()`
Format compliance	`json_validator()`, `regex_match()`
Completeness	`no_refusals()`, `require_tool_use()`
Accuracy	`hallucination_detector()`, `llm_judge()`

For Compliance

Requirement	Recommended Guardrails
GDPR/HIPAA	`pii_detector()`
Content moderation	`toxicity_detector()`
Audit trail	Enable telemetry
Human review	Human-in-the-loop

Performance Considerations

Guardrails have different performance characteristics:

Speed	Guardrails
Fast (under 5ms)	`length_limit()`, `blocked_keywords()`, `rate_limiter()`
Medium (10-50ms)	`pii_detector()`, `prompt_injection()`, `secret_redaction()`, `regex_match()`
Slow (100ms+)	`llm_judge()`, `toxicity_detector()` (ML-based)

Extending Built-in Guardrails

All built-in guardrails can be customized via parameters. For more complex logic, see Custom Guardrails.

Next Steps

Explore individual guardrail documentation in the sidebar
Learn about Custom Guardrails
Set up Logfire Integration for observability