AI Testing Framework: Building a Complete Quality Assurance Pipeline for LLMs and Agents

This article introduces the open-source project ai-testing-prompts-agents developed by Cristian N. The project builds a comprehensive quality assurance pipeline for large language models (LLMs), prompts, and autonomous AI agents. It integrates Promptfoo and DeepEval to enable offline evaluation and visual analysis, helping teams address challenges in LLM output quality, stability, and security—with low cost and no cloud dependency.

With the popularity of LLM and generative AI applications, traditional software testing struggles to handle their random and open-ended outputs. Enterprise-level cloud evaluation services are costly and raise data privacy concerns. Cristian N., a QA engineer with over 20 years of software testing experience, initiated this project to build a complete offline quality assurance pipeline, enabling timely detection of issues like model degradation, prompt drift, and agent behavior anomalies.

The framework adopts a modular design, divided into two core testing areas:

1. Prompt Testing Module (integrated with Promptfoo): Supports prompt matrix evaluation, custom evaluators (enforcing business rule constraints), and guardrail assertion mechanisms (boundary checks);
2. Agent Testing Module (DeepEval and PyTest): Provides LangChain integration demos, zero-cost custom metrics (based on Llama3/Groq), and RAG and answer relevance verification.

The project builds an offline data processing workflow:

• Automated test results are exported as CSV files (eval_results.csv);
• Jupyter Notebook (analysis.ipynb) supports interactive in-depth analysis;
• Streamlit visual dashboard (dashboard.py) provides analysis of pass rates, latency distribution, and failure causes, allowing non-technical personnel to understand model quality.

The project uses a dual-stack architecture: Node.js environment for running Promptfoo prompt tests; Python3.12+ environment for agent evaluation and data analysis. External dependencies recommend using Groq API Key (or OpenAI Key) to ensure evaluation quality while controlling costs.

This pipeline can serve as a continuous integration checkpoint for AI teams. Typical application scenarios include: regression testing before prompt version upgrades, verification of model switching effects, behavior consistency checks after agent workflow changes, and quality gates before production deployment. By tracking empirical metric scores, enterprises can iteratively improve LLM applications with confidence, avoiding performance degradation or hallucination issues.

The ai-testing-prompts-agents project provides a practical and complete open-source solution for AI quality assurance. Its core value is bringing enterprise-level evaluation capabilities to small and medium-sized teams with zero cloud dependency and low cost. It is not just a toolset but also a demonstration of quality-first engineering thinking. In the future, such automated testing infrastructure will become a standard configuration in the industry.