How sarthakrastogi/rival works
Project type and purpose
• This is an open-source Python library and SDK (“Rival AI”) for AI safety in production:
• Real-time attack detection (two built-in detectors: Bhairava & Narasimha)
• Automated red-teaming and benchmarking of AI agents
• Consumers install the package, import detectors/evaluators/generators, and integrate into their own AI pipelines.
File structure
• Top-level:
• setup.py, requirements.txt → Python packaging
• src/rival_ai → core code
• examples/ → usage examples for attack detection & red-teaming
• tests/ → unit and integration tests
• src/rival_ai/core/ → fundamental abstractions: AgentDefinition, Testcase, Benchmarking, Evaluation
• src/rival_ai/detectors/ → plugin detectors (Bhairava, Narasimha), each with data/models/training utilities
• src/rival_ai/generators/ → attack_criteria definitions and testcase_generator
• src/rival_ai/evaluators/ → local_evaluator and prompt templates
• src/rival_ai/utils/ → LLM wrappers, parsing, agent representation
README for additional insights
• Highlights two detection models (Bhairava-0.4B, Narasimha-0.6B)
• Describes “detect_attack” APIs, “detect_attack_with_details” for reasoning
• Mentions red-teaming: automated generation + running of attack scenarios
Detail
• Subcomponents (e.g., detectors/bhairava/models/classifier.py vs. detectors/narasimha/detector.py)
• How config.py drives environment settings
• Example scripts folder as “Usage Examples” block