Responsiblesafeai

Lyzr API Responsible & Safe AI Documentation

1. Introduction

Ensuring Responsible & Safe AI is critical in AI development. Lyzr provides configurable features that help in mitigating biases, preventing toxicity, ensuring factual correctness, and guarding against prompt injection attacks.

2. Responsible AI Features

a) Context Relevance

• Ensures AI responses are aligned with the input query.
• Prevents AI from drifting away from the intended topic.

b) Reflection

• The AI evaluates its own response before generating a final answer.
• Improves logical consistency.

c) Groundedness

• Ensures AI responses are based on verifiable facts.
• Helps maintain accuracy and prevent hallucinations.
• Configuration:
  • Define key facts AI should adhere to.
  • Example:
    {
      "type": "GROUNDEDNESS",
      "config": {
        "facts": [
          "Medical advice should be from verified sources.",
          "AI should not provide emergency medical help."
        ]
      }
    }
    

3. Safe AI Features

a) Fairness & Bias

• Detects and mitigates biased responses.
• Ensures inclusivity and ethical AI behavior.

b) Toxicity Check

• Filters out harmful, offensive, or inappropriate language.
• Helps maintain user-friendly interactions.

c) Prompt Injection Protection

• Prevents users from manipulating AI into behaving undesirably.
• Configuration:
  • Set a threshold (0-1) to determine sensitivity.
  • Example:
    {
      "type": "PROMPT_INJECTION",
      "config": {
        "threshold": 1
      }
    }
    

4. Example Implementation

Here is a JSON configuration example for a Healthcare AI Assistant implementing Responsible & Safe AI features:

{
  "name": "Healthcare AI Assistant",
  "description": "An AI-powered assistant providing medical insights while ensuring responsible and safe AI usage.",
  "agent_role": "You are a trusted medical assistant providing evidence-based health insights.",
  "agent_instructions": "Use medical knowledge and reference trusted sources to provide accurate and responsible healthcare information. Avoid diagnosing or prescribing medications.",
  "examples": null,
  "features": [
    {
      "type": "CONTEXT_RELEVANCE",
      "priority": 0,
      "config": {}
    },
    {
      "type": "SRS",
      "config": {
        "max_tries": 2,
        "modules": {
          "reflection": true,
          "bias": true,
          "toxicity": true
        }
      },
      "priority": 0
    },
    {
      "type": "GROUNDEDNESS",
      "config": {
        "facts": [
          "All medical advice should be referenced from verified healthcare sources.",
          "The assistant should not provide emergency medical advice."
        ]
      },
      "priority": 0
    },
    {
      "type": "PROMPT_INJECTION",
      "config": {
        "threshold": 1
      },
      "priority": 0
    }
  ],
  "tool": null,
  "tool_usage_description": null,
  "provider_id": "OpenAI",
  "temperature": "0.7",
  "top_p": "0.9",
  "llm_credential_id": "lyzr_openai",
  "model": "gpt-4o-mini"
}

5. Conclusion

By leveraging Lyzr’s Responsible & Safe AI features, developers can create AI agents that are ethical, factual, and free from harmful content. Proper configuration of these features enhances trustworthiness and usability while preventing AI manipulation and bias-related issues.