Giskard-AI/giskard

Features

• Model Red-Teaming - Provides automated adversarial testing and vulnerability scanning to identify safety failures and biases in AI models.
• Model Performance Monitoring - Provides a comprehensive system for tracking ML model quality, drift, and reliability via dashboards and scheduling.
• Automated Model Judges - Uses large language models as automated judges to grade the quality and correctness of other model responses.
• RAG Evaluation Dataset Generation - Automatically generates question-answer pairs from knowledge bases to benchmark RAG system accuracy.
• Groundedness Verifications - Verifies that AI responses are strictly based on the provided source context to identify unverified added information.
• Hallucination Detection - Provides automated tools to identify and score hallucinations by verifying if responses are supported by cited sources.
• RAG Pipeline Validation - Measures the groundedness and factual accuracy of retrieval augmented generation systems to detect hallucinations.
• LLM Evaluation Frameworks - Offers a complete framework for measuring language model accuracy and detecting regressions through systematic experiments.
• LLM Testing Libraries - Offers a specialized library for generating test cases and running regression tests to detect hallucinations and bias.
• Factuality Benchmarking Frameworks - Includes frameworks to evaluate the factual consistency of models and their resistance to common misconceptions.
• Model Performance Evaluators - Quantifies the overall reliability and performance of language models by detecting bias, security issues, and performance failures.
• RAG Evaluation Frameworks - Provides a framework for measuring the accuracy and groundedness of retrieval-augmented generation pipelines.
• Omission Detectors - Detects when critical information from the knowledge base is omitted from the generated response using semantic similarity.
• RAG Grounding Verifiers - Verifies if model responses are factually supported by retrieved contexts to detect hallucinations in RAG pipelines.
• Adversarial Input Generation - Generates synthetic edge cases and adversarial inputs to stress-test model resilience and robustness.
• Adversarial Probing Loops - Implements iterative, adaptive prompt-response cycles to intentionally trigger model failures and exploit security weaknesses.
• Security Vulnerability Scanning - Scans for security flaws and business logic failures using automated testing toolkits to ensure system stability.
• AI Prompt Injection Vulnerabilities - Detects security flaws where external inputs manipulate the intended behavior of AI models using safety constraints and string matching.
• LLM Evaluation - Quantifies the accuracy and reliability of LLM agents using automated datasets and specialized quality metrics.
• Test Case Generators - Automatically generates test scenarios and edge cases derived from knowledge bases and conversation logs.
• Report Execution Schedulers - Executes evaluation tests on a fixed timetable to monitor model stability and track performance results over time.
• Agent Boundary Evaluators - Assesses if autonomous agents maintain safety boundaries and avoid harm while executing complex tasks.
• Fairness and Bias Detection Mechanisms - Identifies stereotypes and discriminatory behavior using adversarial datasets and fairness evaluation checks.
• Domain-Specific Reasoning Evaluation - Evaluates model accuracy and reasoning using specialized benchmarks tailored for medical, financial, and legal domains.
• Evaluation Dataset Management - Stores and versions test cases in a centralized repository to maintain consistency across evaluations.
• Function Call Verifiers - Validates the ability to trigger correct functions and APIs across multiple languages, including parallel execution.
• Reasoning Evaluations - Evaluates common sense reasoning and academic knowledge retrieval across diverse logic-based datasets.
• Adversarial Robustness Testing - Quantifies model stability and security by evaluating stress handling via adversarial attack methods.
• Conversational Coherence Metrics - Evaluates the model's ability to maintain context and provide consistent responses across multiple conversational exchanges.
• Model Refusal Detections - Identifies instances where the model incorrectly refuses to answer legitimate queries based on defined rules.
• Semantic Similarity Calculation - Implements vector-based distance metrics to calculate the semantic similarity between generated model responses and reference answers.
• Business Failure Detection - Identifies operational failures in agents using specialized datasets to ensure reliability within a business context.
• Evaluation Schedulers - Automates quality monitoring on daily, weekly, or monthly timetables to track model performance over time.
• Mathematical Reasoning Evaluations - Tests mathematical reasoning by assessing both the final correctness and the quality of the multi-step reasoning process.
• Security Risk Assessments - Evaluates threats and vulnerabilities by testing models against curated security datasets to identify potential exploits.
• AI Output Safety Filters - Provides tools to scan AI-generated content against safety and compliance rules to detect offensive or illegal output.
• AI Content Filters - Generates adversarial test cases to detect violent, illegal, or inappropriate output in model responses.
• Information Disclosure Detection - Generates adversarial test cases to identify the leakage of internal system details or confidential data.
• Private Knowledge Base Imports - Supports uploading document sets in JSON/JSONL formats to build knowledge bases for automated content classification.
• Continuous Monitoring - Monitors for new threats through automated, ongoing security testing and stability tracking.
• Vulnerability Probing Modules - Executes specialized security modules to probe for domain-specific vulnerabilities and systemic weaknesses based on model descriptions.
• Threat-Based Scanning - Performs automated red-teaming based on industry threat categories to identify systemic security risks.
• Enterprise Logic Evaluation - Checks models against enterprise-grade datasets to detect failures in business logic and operational requirements.
• Custom Validation Rules - Allows the definition of specialized failure categories and custom rules to validate domain-specific business requirements.
• Boundary Verification - Generates test datasets and applies validation metrics to verify that models stay within defined business boundaries.
• Evaluation Pipelines - Triggers automated test suites on a fixed timetable to monitor model stability and performance drift.
• Evaluation Dashboards - Ships a visual dashboard to track success rates using correctness and groundedness metrics over time.
• AI Regression Testing Suites - Converts detected vulnerabilities into reusable test suites to prevent performance degradation in future model iterations.
• Code Generation Benchmarks - Benchmarks the ability to solve programming tasks by evaluating function completion and algorithmic correctness.
• Regression Testing Suites - Stores failed test cases as reusable datasets to track performance improvements across model iterations.
• Evaluation and Observability - Testing framework for bias and robustness checks.
• Evaluation Frameworks - Open-source testing and evaluation for machine learning and language models.
• Model Evaluation and Benchmarking - Library for detecting performance, bias, and security issues in AI.
• Observability and Evaluation - Testing framework for detecting bias and errors in models.
• Testing and Observability - Evaluation and testing framework for AI systems.

Star history