optimizer

type AIOptimizer ¶

type AIOptimizer interface {
	// Optimize suggests and applies optimizations for a running engine.
	Optimize(ctx context.Context, id string, e *engine.Engine, status telemetry.StatusUpdate) error
}

AIOptimizer defines the interface for AI-driven performance optimization.

type ConnectFunc ¶ added in v1.18.0

type ConnectFunc func(ctx context.Context, workflowID string) (llm.Provider, string, error)

ConnectFunc returns the provider and model to use for a workflow, resolving its key for that workflow's vhost.

type CorrectionAction ¶

type CorrectionAction string

CorrectionAction defines the type of fix to apply.

const (
	ActionScaleBatch     CorrectionAction = "scale_batch"
	ActionAdjustTimeout  CorrectionAction = "adjust_timeout"
	ActionIncreaseRetry  CorrectionAction = "increase_retry"
	ActionNotifyOperator CorrectionAction = "notify_operator"
	ActionSafeMode       CorrectionAction = "safe_mode"
	ActionSuggestMapping CorrectionAction = "suggest_mapping"
)

type DefaultAIOptimizer ¶

type DefaultAIOptimizer struct {
	// contains filtered or unexported fields
}

DefaultAIOptimizer implements AIOptimizer using the SelfHealingService.

func NewAIOptimizer(aiSvc *ai.SelfHealingService, logger hermod.Logger) *DefaultAIOptimizer

func (a *DefaultAIOptimizer) Optimize(ctx context.Context, id string, e *engine.Engine, status telemetry.StatusUpdate) error

type FailurePattern ¶

type FailurePattern struct {
	ID          string
	Description string
	Threshold   float64
	Window      time.Duration
}

FailurePattern represents a detected issue in the pipeline.

type LLMMappingAdvisor ¶ added in v1.18.0

type LLMMappingAdvisor struct {
	// contains filtered or unexported fields
}

LLMMappingAdvisor asks a model, through pkg/llm, for a mapping. Every string in the sample is run through the PII masker before it is sent.

func NewLLMMappingAdvisor(connect ConnectFunc) *LLMMappingAdvisor

func (a *LLMMappingAdvisor) SuggestMapping(ctx context.Context, workflowID, nodeID string, sample map[string]any) (string, error)

type MappingAdvisor ¶ added in v1.18.0

type MappingAdvisor interface {
	SuggestMapping(ctx context.Context, workflowID, nodeID string, sample map[string]any) (string, error)
}

MappingAdvisor suggests a field mapping for a node whose validation keeps failing, from a sample of what reaches it.

type Optimizer ¶

type Optimizer struct {
	// contains filtered or unexported fields
}

Optimizer tracks engine metrics and suggests/applies runtime tunings.

func NewOptimizer(logger hermod.Logger, aiOpt AIOptimizer, notifier func(workflowID, title, message string)) *Optimizer

func (o *Optimizer) Register(id string, e *engine.Engine)

func (o *Optimizer) SetMappingAdvisor(a MappingAdvisor)

func (o *Optimizer) SetProposer(p Proposer)

func (o *Optimizer) Start(ctx context.Context)

func (o *Optimizer) Unregister(id string)

type OptimizerConfig ¶

type OptimizerConfig struct {
	OptimizationInterval time.Duration
	EnableAI             bool
}

type Proposer ¶ added in v1.18.0

type Proposer interface {
	Propose(ctx context.Context, s Suggestion) error
}

Proposer keeps a suggestion for approval.

type SelfCorrectionGate ¶

type SelfCorrectionGate struct {
	// contains filtered or unexported fields
}

SelfCorrectionGate watches a workflow's telemetry for failure patterns.

It never changes a running engine or a stored workflow. A fix becomes a Suggestion handed to the Proposer, which keeps it for a person to approve; everything else is advice to the operator.

func NewSelfCorrectionGate(logger hermod.Logger, notifier func(workflowID, title, message string)) *SelfCorrectionGate

func (g *SelfCorrectionGate) Analyze(id string, _ *engine.Engine, status telemetry.StatusUpdate)

func (g *SelfCorrectionGate) SetMappingAdvisor(a MappingAdvisor)

func (g *SelfCorrectionGate) SetProposer(p Proposer)

type Suggestion ¶ added in v1.18.0

type Suggestion struct {
	WorkflowID string
	NodeID     string
	Action     CorrectionAction
	Reason     string
}

Suggestion is a fix the gate would make. It is not applied: a Proposer stores it for a person to approve, and only an approval changes the workflow, through the same path as an edit in the editor.