Skip to main content

Class: DualLLMAuditor

Defined in: security/DualLLMAuditor.ts:79

Dual-LLM Auditor implementing IGuardrailService.

This is the second layer of defense in the Wunderland security pipeline. It uses a separate, fast model to audit outputs from the primary model.

Key benefits:

  • Uncensored primary models can be audited by a more constrained model
  • Fast 3B model provides real-time audit without significant latency
  • Independent evaluation prevents prompt injection from compromising both models

Example​

const auditor = new DualLLMAuditor({
  auditorProviderId: 'ollama',
  auditorModelId: 'llama3.2:3b',
  evaluateStreamingChunks: true,
}, providerManager);

// Register as a guardrail
orchestrator.registerGuardrail(auditor);

Implements​

Constructors​

Constructor​

new DualLLMAuditor(config?, invokeAuditor?): DualLLMAuditor

Defined in: security/DualLLMAuditor.ts:91

Parameters​

config?​

Partial<DualLLMAuditorConfig> = {}

invokeAuditor?​

(prompt) => Promise<string>

Returns​

DualLLMAuditor

Properties​

config​

readonly config: GuardrailConfig

Defined in: security/DualLLMAuditor.ts:80

Methods​

evaluateInput()​

evaluateInput(payload): Promise<any>

Defined in: security/DualLLMAuditor.ts:117

Captures the user request for context in output evaluation.

Parameters​

payload​

GuardrailInputPayload

Returns​

Promise<any>

evaluateOutput()​

evaluateOutput(payload): Promise<any>

Defined in: security/DualLLMAuditor.ts:132

Audits output chunks from the primary model.

Parameters​

payload​

GuardrailOutputPayload

Returns​

Promise<any>

getConfig()​

getConfig(): DualLLMAuditorConfig

Defined in: security/DualLLMAuditor.ts:387

Gets the current auditor configuration.

Returns​

DualLLMAuditorConfig

performAudit()​

performAudit(userRequest, assistantResponse): Promise<AuditResult>

Defined in: security/DualLLMAuditor.ts:187

Performs the actual audit using the auditor model.

Parameters​

userRequest​

string

assistantResponse​

string

Returns​

Promise<AuditResult>

resetEvaluationCount()​

resetEvaluationCount(): void

Defined in: security/DualLLMAuditor.ts:379

Resets the evaluation count (call at start of new request).

Returns​

void