Documentation
¶
Overview ¶
Package guardrails 提供 Agent 输出质量护栏
与 security/guard 不同，guardrails 关注的是输出质量而非安全性：
- 函数式检查：长度、关键词、格式、自定义规则
- LLM-as-Judge：使用 LLM 评估输出质量
- 幻觉检测：检测输出是否包含虚构信息
- 相关性检查：检测输出是否与输入相关
使用示例：
rail := NewGuardrail(
WithLengthCheck(50, 2000),
WithKeywordCheck([]string{"错误", "无法"}),
WithLLMJudge(llmProvider, "model-name"),
)
result, err := rail.Check(ctx, input, output)
if !result.Passed {
// 输出质量不合格，需要重新生成
}
Index ¶
- type CheckDetail
- type CheckResult
- type Checker
- type Guardrail
- type GuardrailOption
- func WithChecker(checker Checker) GuardrailOption
- func WithFailFast(failFast bool) GuardrailOption
- func WithFormatCheck(format OutputFormat) GuardrailOption
- func WithHallucinationCheck(provider llm.Provider, model string, context string) GuardrailOption
- func WithKeywordCheck(forbidden []string) GuardrailOption
- func WithLLMJudge(provider llm.Provider, model string) GuardrailOption
- func WithLengthCheck(minLen, maxLen int) GuardrailOption
- func WithRelevanceCheck(provider llm.Provider, model string) GuardrailOption
- func WithThreshold(threshold float64) GuardrailOption
- type OutputFormat
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type CheckDetail ¶
type CheckDetail struct {
// Name 检查项名称
Name string `json:"name"`
// Passed 是否通过
Passed bool `json:"passed"`
// Score 分数
Score float64 `json:"score"`
// Message 说明信息
Message string `json:"message,omitempty"`
}
CheckDetail 单项检查详情
type CheckResult ¶
type CheckResult struct {
// Passed 是否通过
Passed bool `json:"passed"`
// Score 质量分数 (0-1，越高质量越好)
Score float64 `json:"score"`
// Reason 不通过的原因
Reason string `json:"reason,omitempty"`
// Suggestions 改进建议
Suggestions []string `json:"suggestions,omitempty"`
// Details 详细检查结果
Details []CheckDetail `json:"details,omitempty"`
}
CheckResult 检查结果
type Checker ¶
type Checker interface {
// Name 检查器名称
Name() string
// Check 执行检查
// input 是用户输入，output 是 Agent 输出
Check(ctx context.Context, input, output string) (*CheckDetail, error)
}
Checker 单项检查器接口
type Guardrail ¶
type Guardrail struct {
// contains filtered or unexported fields
}
Guardrail 输出护栏 组合多个 Checker 对 Agent 输出进行质量检查
type GuardrailOption ¶
type GuardrailOption func(*Guardrail)
GuardrailOption 护栏配置选项
func WithFormatCheck ¶
func WithFormatCheck(format OutputFormat) GuardrailOption
WithFormatCheck 添加格式检查
func WithHallucinationCheck ¶
func WithHallucinationCheck(provider llm.Provider, model string, context string) GuardrailOption
WithHallucinationCheck 添加幻觉检测
func WithKeywordCheck ¶
func WithKeywordCheck(forbidden []string) GuardrailOption
WithKeywordCheck 添加关键词检查（输出不应包含这些关键词）
func WithLLMJudge ¶
func WithLLMJudge(provider llm.Provider, model string) GuardrailOption
WithLLMJudge 添加 LLM-as-Judge 检查
func WithLengthCheck ¶
func WithLengthCheck(minLen, maxLen int) GuardrailOption
WithLengthCheck 添加长度检查
func WithRelevanceCheck ¶
func WithRelevanceCheck(provider llm.Provider, model string) GuardrailOption
WithRelevanceCheck 添加相关性检查
func WithThreshold ¶
func WithThreshold(threshold float64) GuardrailOption
WithThreshold 设置通过阈值（0-1）
type OutputFormat ¶
type OutputFormat int
OutputFormat 输出格式类型
const ( // FormatJSON JSON 格式 FormatJSON OutputFormat = iota // FormatMarkdown Markdown 格式 FormatMarkdown // FormatPlainText 纯文本格式 FormatPlainText )
Source Files
Click to show internal directories.
Click to hide internal directories.