Infosys/Infosys-Responsible-AI-Toolkit

What it solves

This toolkit addresses the need for trustworthiness and transparency in AI solutions. It provides a comprehensive set of APIs to integrate safety, security, privacy, explainability, fairness, and hallucination detection into both Generative AI and traditional Machine Learning models, helping developers regulate content and mitigate risks.

How it works

The toolkit is organized into modular APIs and a supporting user interface. It uses a "Moderation Layer" to regulate LLM prompts and responses, and specialized APIs for different tenets of responsible AI:

• Safety & Privacy: Detects and anonymizes PII, toxic text, and profane content.
• Explainability: Provides global and local explanations for ML models (using SHAP and LIME) and specific methods for LLMs (such as Chain of Thought and Graph of Thought).
• Fairness: Detects biases using statistical methods and provides mitigation techniques like re-weighing.
• Hallucination: Quantifies hallucinations in LLM responses, specifically within RAG scenarios.
• Security: Checks for prompt injections and jailbreaks.

Who it’s for

Developers and AI practitioners who need to implement guardrails and ethical alignment for AI applications, particularly those optimized for Azure OpenAI.

Highlights

• Comprehensive Guardrails: Includes checks for prompt injection, jailbreaks, toxicity, and restricted topics.
• RAG-Specific Detection: Dedicated API for quantifying hallucinations in Retrieval-Augmented Generation scenarios.
• Model Transparency: Supports multiple explainability frameworks including Thread of Thoughts (ThoT) and Chain of Verification (CoVe).
• Full-Stack Toolkit: Includes a micro-frontend UI (Angular), a Python backend, and telemetry integration with Elasticsearch and Kibana.