LLM Evaluation & Testing

关于

This skill streamlines the process of evaluating Large Language Model (LLM) outputs by integrating the Promptfoo framework directly into the development workflow. It enables developers to perform automated regression testing, compare prompt versions, and conduct security red-teaming to detect jailbreaks or PII leaks. By providing a structured way to implement CI/CD quality gates, it ensures that only high-quality, safe, and cost-effective prompts are deployed to production, moving beyond subjective evaluations to objective, data-driven metrics for RAG systems and standalone prompts.

主要功能

• Seamless CI/CD integration with GitHub Actions and quality gates
• Security-focused red teaming for jailbreak and PII leak detection
• Side-by-side model and prompt performance comparisons
• Automated regression testing for LLM prompts and RAG systems
• 2 GitHub stars
• Semantic assertions using LLM-as-a-judge and factuality rubrics

使用场景

• Setting up automated quality gates in CI/CD pipelines to block degraded prompt performance
• Comparing the cost, latency, and accuracy of different models like Claude 3.5 and GPT-4o
• Performing security audits on chatbots to identify potential prompt injection vulnerabilities