Skip to main content

🛡️ Building Responsible AI Apps with AWS Guardrails

In the world of Retrieval-Augmented Generation (RAG), context validation and content safety are non-negotiable—especially when your app is dealing with sensitive or regulated information.

In this post, we'll explore how to use AWS Guardrails, a fully managed service in Amazon Bedrock, to validate prompts and responses in RAG applications and enforce responsible AI policies at scale.

👉 AWS Guardrails Official Docs

🤖 Why Guardrails Matter in RAG Applications​

When building LLM apps—especially RAG-based ones—you need strong controls over:

  • ❌ Prompt Injection risks
  • 🧠 Hallucinated responses from foundation models
  • ⚠️ Unsafe content, including harmful or biased language
  • 🔒 Sensitive data exposure (like PII)

AWS Guardrails acts as a contextual safety layer between user inputs and model outputs, helping you monitor, block, or redact unsafe, off-topic, or ungrounded interactions — in both prompts and responses.

🧩 Guardrails Capabilities Overview​

Guardrails is fully configurable with granular policies for:

Policy TypePurpose
topicPolicyConfigBlock specific subject areas (e.g., investment advice)
contentPolicyConfigDetect and moderate harmful content (e.g., hate, violence, misconduct)
wordPolicyConfigFlag or block specific keywords or phrases
sensitiveInformationPolicyConfigAnonymize or block PII and sensitive data
contextualGroundingPolicyConfigSet thresholds for relevance and factual grounding
blockedInputMessagingDefine fallback messages for blocked prompts
blockedOutputsMessagingDefine fallback messages for blocked responses

📐 Architecture with Guardrails​

Here’s a simplified architecture of a RAG chatbot using AWS Bedrock + Guardrails:

Architecture Diagram

🛠️ Sample Guardrails Configuration (Python Snippet)​

description = "Prevents our model from providing proprietary information."

topicPolicyConfig = {
    "topicsConfig": [
        {
            "name": "AI Social Journal Denied Topics",
            "definition": "Personalized financial recommendations or fiduciary advice.",
            "examples": [
                "What stocks should I invest in?",
                "How should I allocate my 401(k)?",
            ],
            "type": "DENY"
        }
    ]
}

contentPolicyConfig = {
    "filtersConfig": [
        {"type": "SEXUAL", "inputStrength": "HIGH", "outputStrength": "HIGH"},
        {"type": "VIOLENCE", "inputStrength": "HIGH", "outputStrength": "HIGH"},
        {"type": "HATE", "inputStrength": "HIGH", "outputStrength": "HIGH"},
        {"type": "INSULTS", "inputStrength": "HIGH", "outputStrength": "HIGH"},
        {"type": "MISCONDUCT", "inputStrength": "HIGH", "outputStrength": "HIGH"},
        {"type": "PROMPT_ATTACK", "inputStrength": "HIGH", "outputStrength": "NONE"},
    ]
}

wordPolicyConfig = {
    "wordsConfig": [
        {"text": "financial planning guidance"},
        {"text": "retirement fund suggestions"},
        {"text": "trust fund setup"},
    ],
    "managedWordListsConfig": [{"type": "PROFANITY"}],
}

sensitiveInformationPolicyConfig = {
    "piiEntitiesConfig": [
        {"type": "EMAIL", "action": "ANONYMIZE"},
        {"type": "NAME", "action": "ANONYMIZE"},
        {"type": "US_SOCIAL_SECURITY_NUMBER", "action": "BLOCK"},
    ],
    "regexesConfig": [
        {
            "name": "Account Number",
            "description": "Matches 10-digit account numbers",
            "pattern": r"\b\d{6}\d{4}\b",
            "action": "ANONYMIZE"
        }
    ]
}

contextualGroundingPolicyConfig = {
    "filtersConfig": [
        {"type": "GROUNDING", "threshold": 0.75},
        {"type": "RELEVANCE", "threshold": 0.75}
    ]
}

blockedInputMessaging = "Sorry, I can't assist with that."
blockedOutputsMessaging = "Sorry, I can't assist with that."

🚀 Live Demo with Streamlit​

To demonstrate this setup, I built a Streamlit app that validates both user input and model responses using AWS Guardrails. The app lets you:

  • Ask questions via a chat interface
  • Watch Guardrails intercept unsafe prompts or outputs

🧪 Response from Guardrails​

Here’s how the app behaves:

  • User prompt violates topic policy → receives fallback message
  • Model hallucination exceeds relevance threshold → blocked
  • PII detected in user prompt → redacted automatically

✅ Safety, groundedness, and compliance—by design.

Sample Output with Guardrails

Sample Outpout with regular response

📁 Full Codebase​

You can find the complete working example and integration code on GitHub:
🔗 View on GitHub

🧭 Final Thoughts​

If you’re building production-grade RAG or chatbot applications, AWS Guardrails should be part of your stack from day one.

It helps you:

  • Catch unsafe prompts before they reach the LLM
  • Block hallucinated or off-topic answers
  • Protect sensitive information
  • Build trust in your AI systems

And best of all — it’s configurable, scalable, and fully managed.

🔗 Resources​