Documentation
¶
Overview ¶
Package guardrail provides output validation guardrails for AI agent responses.
Guardrails enforce contracts on agent output — reject, retry, or constrain responses based on structural, lexical, or semantic rules.
Index ¶
- type BannedWordsGuardrail
- type Chain
- type ChainOption
- type ContainsGuardrail
- type ContextGuardrail
- type Guardrail
- type HallucinationGuardrail
- type HallucinationGuardrailOption
- type JSONSchemaGuardrail
- type LLMGuardrail
- type LLMGuardrailOption
- type LengthGuardrail
- type RegexGuardrail
- type ValidationError
- type ValidationFailure
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type BannedWordsGuardrail ¶
type BannedWordsGuardrail struct {
// contains filtered or unexported fields
}
BannedWordsGuardrail validates that output does not contain any banned words.
func NewBannedWordsGuardrail ¶
func NewBannedWordsGuardrail(words ...string) *BannedWordsGuardrail
NewBannedWordsGuardrail creates a BannedWordsGuardrail with the given banned words.
func (*BannedWordsGuardrail) Validate ¶
func (g *BannedWordsGuardrail) Validate(output string) error
Validate checks the output for banned words.
type Chain ¶ added in v0.9.0
type Chain struct {
// contains filtered or unexported fields
}
Chain runs multiple guardrails sequentially and collects all errors. It implements both Guardrail and ContextGuardrail.
func NewChain ¶ added in v0.9.0
func NewChain(guardrails []Guardrail, opts ...ChainOption) *Chain
NewChain creates a Chain wrapping the given guardrails.
type ChainOption ¶ added in v0.9.0
type ChainOption func(*Chain)
ChainOption configures a Chain.
func WithMaxErrors ¶ added in v0.9.0
func WithMaxErrors(n int) ChainOption
WithMaxErrors limits the number of errors collected before stopping (0 = unlimited).
func WithStopOnFirst ¶ added in v0.9.0
func WithStopOnFirst(stop bool) ChainOption
WithStopOnFirst configures whether the chain stops on the first error (default false).
type ContainsGuardrail ¶
type ContainsGuardrail struct {
// contains filtered or unexported fields
}
ContainsGuardrail validates that output contains all required substrings.
func NewContainsGuardrail ¶
func NewContainsGuardrail(substrings ...string) *ContainsGuardrail
NewContainsGuardrail creates a ContainsGuardrail requiring the given substrings.
func (*ContainsGuardrail) Validate ¶
func (g *ContainsGuardrail) Validate(output string) error
Validate checks the output contains all required substrings.
type ContextGuardrail ¶ added in v0.9.0
type ContextGuardrail interface {
Guardrail
ValidateWithContext(ctx context.Context, output string) error
}
ContextGuardrail is an optional extension of Guardrail for guardrails that need runtime context such as task description, agent goal, or source materials. The validation loop detects ContextGuardrail via type assertion and calls ValidateWithContext instead of Validate when available.
type HallucinationGuardrail ¶ added in v0.9.0
type HallucinationGuardrail struct {
// contains filtered or unexported fields
}
HallucinationGuardrail cross-checks output claims against source context using an LLM.
func NewHallucinationGuardrail ¶ added in v0.9.0
func NewHallucinationGuardrail(sourceContext string, client llm.Client, opts ...HallucinationGuardrailOption) *HallucinationGuardrail
NewHallucinationGuardrail creates a guardrail that detects hallucinations by cross-checking output against the given source context.
func (*HallucinationGuardrail) Validate ¶ added in v0.9.0
func (g *HallucinationGuardrail) Validate(output string) error
Validate checks the output for unsupported claims against the source context.
type HallucinationGuardrailOption ¶ added in v0.9.0
type HallucinationGuardrailOption func(*HallucinationGuardrail)
HallucinationGuardrailOption configures a HallucinationGuardrail.
func WithHallucinationJudgePrompt ¶ added in v0.9.0
func WithHallucinationJudgePrompt(tmpl string) HallucinationGuardrailOption
WithHallucinationJudgePrompt sets a custom judge prompt template with {source_context} and {output} placeholders.
type JSONSchemaGuardrail ¶
type JSONSchemaGuardrail struct {
// contains filtered or unexported fields
}
JSONSchemaGuardrail validates that output is valid JSON matching a JSON Schema (draft-07).
func NewJSONSchemaGuardrail ¶
func NewJSONSchemaGuardrail(schemaStr string) (*JSONSchemaGuardrail, error)
NewJSONSchemaGuardrail creates a JSONSchemaGuardrail with the given JSON Schema string. Returns error if the schema is invalid.
func (*JSONSchemaGuardrail) Validate ¶
func (g *JSONSchemaGuardrail) Validate(output string) error
Validate checks the output conforms to the JSON schema.
type LLMGuardrail ¶ added in v0.9.0
type LLMGuardrail struct {
// contains filtered or unexported fields
}
LLMGuardrail uses an LLM to judge whether the output satisfies a description.
func NewLLMGuardrail ¶ added in v0.9.0
func NewLLMGuardrail(description string, client llm.Client, opts ...LLMGuardrailOption) *LLMGuardrail
NewLLMGuardrail creates an LLMGuardrail with the given description and LLM client.
func (*LLMGuardrail) Validate ¶ added in v0.9.0
func (g *LLMGuardrail) Validate(output string) error
Validate checks whether the output satisfies the guardrail description using the LLM.
type LLMGuardrailOption ¶ added in v0.9.0
type LLMGuardrailOption func(*LLMGuardrail)
LLMGuardrailOption configures an LLMGuardrail.
func WithJudgePrompt ¶ added in v0.9.0
func WithJudgePrompt(tmpl string) LLMGuardrailOption
WithJudgePrompt sets a custom judge prompt template with {output} and {description} placeholders.
type LengthGuardrail ¶
type LengthGuardrail struct {
// contains filtered or unexported fields
}
LengthGuardrail validates that output length is within the specified bounds.
func NewLengthGuardrail ¶
func NewLengthGuardrail(min, max int) *LengthGuardrail
NewLengthGuardrail creates a LengthGuardrail with the given min/max bounds.
func (*LengthGuardrail) Validate ¶
func (g *LengthGuardrail) Validate(output string) error
Validate checks the output length is within bounds.
type RegexGuardrail ¶
type RegexGuardrail struct {
// contains filtered or unexported fields
}
RegexGuardrail validates that output matches a given regular expression.
func NewRegexGuardrail ¶
func NewRegexGuardrail(pattern string) (*RegexGuardrail, error)
NewRegexGuardrail creates a RegexGuardrail. Returns error if pattern is invalid.
func (*RegexGuardrail) Validate ¶
func (g *RegexGuardrail) Validate(output string) error
Validate checks the output against the regex pattern.
type ValidationError ¶ added in v0.9.0
type ValidationError struct {
Errors []ValidationFailure
}
ValidationError aggregates zero or more ValidationFailure entries.
func (*ValidationError) Error ¶ added in v0.9.0
func (e *ValidationError) Error() string
func (*ValidationError) Unwrap ¶ added in v0.9.0
func (e *ValidationError) Unwrap() error
Unwrap returns the first ValidationFailure, enabling errors.Is and errors.As traversal.
type ValidationFailure ¶ added in v0.9.0
ValidationFailure represents a single validation failure with optional field-level metadata.
func (ValidationFailure) Error ¶ added in v0.9.0
func (f ValidationFailure) Error() string
Source Files