防护措施

使用 adk-guardrail 进行输入/输出验证和内容安全处理。

概述

防护措施会验证并转换 agent 的输入和输出，以确保安全性、合规性和质量。它们与 agent 执行并行运行，并可以：

• 阻止有害或无关内容
• 编辑个人身份信息（电子邮箱、电话号码、SSN、信用卡号）
• 对输出强制执行 JSON schema
• 限制内容长度

安装

[dependencies]
adk-guardrail = "2.1.0"

# For JSON schema validation
adk-guardrail = { version = "2.1.0", features = ["schema"] }

核心概念

GuardrailResult

每个防护措施都会返回三种结果之一：

pub enum GuardrailResult {
    Pass,                                    // Content is valid
    Fail { reason: String, severity: Severity },  // Content rejected
    Transform { new_content: Content, reason: String },  // Content modified
}

严重级别

pub enum Severity {
    Low,      // Warning only, doesn't block
    Medium,   // Blocks but continues other checks
    High,     // Blocks immediately
    Critical, // Blocks and fails fast
}

个人身份信息编辑

自动检测并编辑个人身份信息：

use adk_guardrail::{PiiRedactor, PiiType};

// Default: emails, phones, SSNs, credit cards
let redactor = PiiRedactor::new();

// Or select specific types
let redactor = PiiRedactor::with_types(&[
    PiiType::Email,
    PiiType::Phone,
]);

// Direct redaction
let (redacted, found_types) = redactor.redact("Email: test@example.com");
// redacted = "Email: [EMAIL REDACTED]"
// found_types = [PiiType::Email]

支持的个人身份信息类型：

内容过滤

阻止有害内容或强制执行主题约束：

use adk_guardrail::ContentFilter;

// Block harmful content patterns (developer-friendly: excludes "hack"/"exploit")
let filter = ContentFilter::harmful_content();

// Strict variant that also blocks "hack" and "exploit"
let filter = ContentFilter::harmful_content_strict();

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

// Enforce topic relevance
let filter = ContentFilter::on_topic("cooking", vec![
    "recipe".into(),
    "cook".into(),
    "bake".into(),
]);

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

自定义内容过滤器

use adk_guardrail::{ContentFilter, ContentFilterConfig, Severity};

let config = ContentFilterConfig {
    blocked_keywords: vec!["spam".into()],
    required_topics: vec!["rust".into(), "programming".into()],
    max_length: Some(5000),
    min_length: Some(10),
    severity: Severity::High,
};

let filter = ContentFilter::new("custom_filter", config);

模式验证

对代理输出强制执行 JSON 模式（需要 schema 功能）：

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

let schema = json!({
    "type": "object",
    "properties": {
        "name": { "type": "string" },
        "age": { "type": "integer", "minimum": 0 }
    },
    "required": ["name"]
});

let validator = SchemaValidator::new(&schema)?
    .with_name("user_schema")
    .with_severity(Severity::High);

验证器从以下内容中提取 JSON：

• 原始 JSON 文本
• Markdown 代码块（```json ... ```）

GuardrailSet

组合多个防护栏：

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

let guardrails = GuardrailSet::new()
    .with(ContentFilter::harmful_content())
    .with(ContentFilter::max_length(5000))
    .with(PiiRedactor::new());

GuardrailExecutor

运行防护栏并获取详细结果：

use adk_guardrail::{GuardrailExecutor, GuardrailSet, PiiRedactor};
use adk_core::Content;

let guardrails = GuardrailSet::new()
    .with(PiiRedactor::new());

let content = Content::new("user")
    .with_text("Contact: test@example.com");

let result = GuardrailExecutor::run(&guardrails, &content).await?;

if result.passed {
    // Use transformed content if available
    let final_content = result.transformed_content.unwrap_or(content);
    println!("Content passed validation");
} else {
    for (name, reason, severity) in &result.failures {
        println!("Guardrail '{}' failed: {} ({:?})", name, reason, severity);
    }
}

ExecutionResult

pub struct ExecutionResult {
    pub passed: bool,                              // Overall pass/fail
    pub transformed_content: Option<Content>,      // Modified content (if any)
    pub failures: Vec<(String, String, Severity)>, // (name, reason, severity)
}

自定义防护栏

实现 Guardrail trait：

use adk_guardrail::{Guardrail, GuardrailResult, Severity};
use adk_core::Content;
use async_trait::async_trait;

pub struct ProfanityFilter {
    words: Vec<String>,
}

#[async_trait]
impl Guardrail for ProfanityFilter {
    fn name(&self) -> &str {
        "profanity_filter"
    }

    async fn validate(&self, content: &Content) -> GuardrailResult {
        let text: String = content.parts
            .iter()
            .filter_map(|p| p.text())
            .collect();

        for word in &self.words {
            if text.to_lowercase().contains(word) {
                return GuardrailResult::Fail {
                    reason: format!("Contains profanity: {}", word),
                    severity: Severity::High,
                };
            }
        }

        GuardrailResult::Pass
    }

    // Run in parallel with other guardrails (default: true)
    fn run_parallel(&self) -> bool {
        true
    }

    // Fail fast on this guardrail's failure (default: true)
    fn fail_fast(&self) -> bool {
        true
    }
}

与代理集成

防护栏与 LlmAgentBuilder 集成：

配置的输入防护栏会在模型执行前运行，配置的输出防护栏会在生成的响应返回给调用方之前运行。

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

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

let output_guardrails = GuardrailSet::new()
    .with(SchemaValidator::new(&output_schema)?);

let agent = LlmAgentBuilder::new("assistant")
    .model(model)
    .instruction("You are a helpful assistant.")
    .input_guardrails(input_guardrails)
    .output_guardrails(output_guardrails)
    .build()?;

执行流程

User Input
    │
    ▼
┌─────────────────────┐
│  Input Guardrails   │ ← PII redaction, content filtering
│  (parallel)         │
└─────────────────────┘
    │
    ▼ (transformed or blocked)
┌─────────────────────┐
│  Agent Execution    │
└─────────────────────┘
    │
    ▼
┌─────────────────────┐
│  Output Guardrails  │ ← Schema validation, safety checks
│  (parallel)         │
└─────────────────────┘
    │
    ▼
Final Response

示例

cargo check -p adk-guardrail
cargo check -p adk-rust --no-default-features --features guardrail

最佳实践

---

上一页：← 访问控制 | 下一页：记忆 →