SDK & Libraries
Documentation

Guardian SDK: Custom Guardrail Development Kit

Learn how to build, test, and register custom AI governance and audit modules using the Python-based HardRAG SDK.

The Guardian SDK is the definitive framework for building, testing, and deploying custom governance modules within the HardRAG ecosystem. Optimized for Sovereign and Cloud-Native workloads.


🏗️ Core Architecture

The SDK leverages DSPy for programmatic prompt optimization and Pydantic for structured validation. Every Guardian module follows a strict "Consensus Intelligence" signature.

1. Defining a Custom Guardrail

To build a specialized auditor (e.g., Legal Document PII or Banking Transaction Audit), inherit from the base HardRAGModule.

from app.hardrag.modules import HardRAGGuard
import dspy

class FinancialComplianceGuard(dspy.Module):
    """
    Audits financial instructions against internal AML/KYC policies.
    """
    def __init__(self):
        super().__init__()
        self.auditor = dspy.ChainOfThought("prompt, policy -> verdict, rationale")

    def forward(self, prompt, policy):
        # The 'Sovereign Consensus' logic
        result = self.auditor(prompt=prompt, policy=policy)
        return dspy.Prediction(verdict=result.verdict, rationale=result.rationale)

2. Integration with HardRAG Guard

Once your module is defined, register it with the HardRAGGuard orchestrator to enable multi-agent consensus.

from app.hardrag.guard import HardRAGGuard

# Initialize the Sovereign Guard
guard = HardRAGGuard()

# Add your custom Guardian module
guard.register_module(
    id="aml_auditor",
    module=FinancialComplianceGuard(),
    tier="PLATFORM"
)

🌀 Continuous Self-Improvement

Guardian SDK modules are natively compatible with the HardRAG Feedback Loop.

  • Training: Captured audit logs are automatically used as "Few-Shot" examples.
  • Optimization: Use optimizer.run(module) to refine prompt weights based on CISO-corrected history.

🔒 Security & Performance

  • Zero-Knowledge Logs: All training data is sanitized before entering the optimization loop.
  • Cryptographic Audit Integrity: Modules registered at the PLATFORM tier automatically receive a SHA-256-SIGNED integrity seal in the audit feed.
  • Latency Target: Optimal Guardian modules should maintain an overhead of < 150ms.

🚀 Deployment (Sovereign Mode)

To deploy your custom guardrails in an air-gapped environment:

  1. Compile your DSPy program to optimized_weights.json.
  2. Map the weights directory to the Sovereign Vault container.
  3. HardRAG will detect the custom signatures and activate the Certified Trace.

[!TIP] Developer Support: For partners, the HardRAG engineering team provides technical support for complex module development.