adk-guardrail

Guardrails framework for ADK agents - input/output validation, content filtering, PII redaction.

Overview

adk-guardrail provides safety and validation for AI agents in the ADK-Rust framework:

• PiiRedactor - Detect and redact PII (Email, Phone, SSN, Credit Card, IP Address)
• ContentFilter - Block harmful content, off-topic responses, max length
• SchemaValidator - Validate JSON output against schemas
• GuardrailSet - Compose multiple guardrails with parallel execution

Installation

[dependencies]
adk-guardrail = "0.3.2"

# With JSON schema validation (default)
adk-guardrail = { version = "0.3.2", features = ["schema"] }

Quick Start

PII Redaction

use adk_guardrail::{PiiRedactor, Guardrail};
use adk_core::Content;

let redactor = PiiRedactor::new();
let content = Content::new("user").with_text("Contact me at test@example.com");
let result = redactor.validate(&content).await;

// Result is Transform with "[EMAIL REDACTED]"

Content Filtering

use adk_guardrail::ContentFilter;

// Block harmful content
let filter = ContentFilter::harmful_content();

// Keep responses on-topic
let filter = ContentFilter::on_topic("cooking", vec!["recipe".into(), "bake".into()]);

// Limit response length
let filter = ContentFilter::max_length(1000);

// Block specific keywords
let filter = ContentFilter::blocked_keywords(vec!["forbidden".into()]);

Agent Integration

Requires guardrails feature on adk-agent:

adk-agent = { version = "0.3.2", features = ["guardrails"] }

use adk_agent::LlmAgentBuilder;
use adk_guardrail::{GuardrailSet, ContentFilter, PiiRedactor};

let input_guardrails = GuardrailSet::new()
    .with(ContentFilter::harmful_content())
    .with(PiiRedactor::new());

let agent = LlmAgentBuilder::new("assistant")
    .input_guardrails(input_guardrails)
    .build()?;

Guardrail Results

Each guardrail returns one of three results:

Severity Levels

Built-in Guardrails

PiiRedactor

Detects and redacts personally identifiable information:

• Email → [EMAIL REDACTED]
• Phone → [PHONE REDACTED]
• Ssn → [SSN REDACTED]
• CreditCard → [CREDIT CARD REDACTED]
• IpAddress → [IP REDACTED]

ContentFilter

Validates content against rules:

• harmful_content() - Blocks common harmful patterns
• on_topic(topic, keywords) - Ensures topic relevance
• max_length(n) - Limits content length
• blocked_keywords(list) - Blocks specific words

SchemaValidator

Validates JSON output against a schema (requires schema feature):

use adk_guardrail::SchemaValidator;
use serde_json::json;

let schema = json!({
    "type": "object",
    "properties": {
        "name": { "type": "string" },
        "age": { "type": "integer" }
    },
    "required": ["name"]
});

let validator = SchemaValidator::new("user_schema", schema)?;

Features

• Parallel guardrail execution with early exit on failure
• Composable with GuardrailSet builder pattern
• Integration with LlmAgentBuilder via input_guardrails() / output_guardrails()
• Async validation with the Guardrail trait

Related Crates

• adk-rust - Meta-crate with all components
• adk-agent - Agent implementations with guardrail support
• adk-core - Core traits and types

License

Apache-2.0

Part of ADK-Rust

This crate is part of the ADK-Rust framework for building AI agents in Rust.