Vera – AI Feature Testing Engine

**Vera ** is an extensible testing engine designed to bring software engineering rigor to AI feature development. It provides a standardized framework for evaluating Generative AI outputs using a **Hybrid Evaluation ** approach: combining deterministic static checks with semantic "LLM-as-a-Judge" evaluation.

Why Vera?

Building reliable AI features is hard. Traditional unit tests struggle with the probabilistic nature of LLMs, while manual testing is unscalable and subjective. Vera bridges this gap.

• **🤖 Hybrid Evaluation **: strict programmatic validation (e.g., "is valid JSON?") combined with nuanced AI grading (e.g., "is the tone professional?").
• **📏 Spec-Driven Quality **: Define success using natural language Rubrics, Safety Constraints, and Golden Datasets that the LLM Judge follows.
• ⚡ High Performance: Built on asyncio and anyio for parallel test execution.
• 🔌 Plugin Architecture: deeply extensible via pluggy. Add new features, commands, or reporters easily.
• **📊 Standardized Reporting **: Get detailed CSV reports with granular scores, reasoning, and pass/fail status.

Documentation

• Installation and Usage: Setup guide and CLI reference.
• Testing Philosophy: Deep dive into "LLM-as-a-Judge" and Test Specs.
• Plugin Development: How to build custom evaluation plugins.
• Contributing: Guide for contributing to the Vera core engine.

How It Works

Vera evaluates your AI feature in four steps:

1. Execution: Runs your feature against a set of inputs (Test Cases).
2. Static Analysis: Runs Python-based checks on the output (e.g., regex, syntax validation).
3. LLM Evaluation: An "LLM Judge" (e.g., Gemini) grades the output based on your provided * Specs* (Rubrics, Safety Constraints, Style Guides).
4. Reporting: Aggregates all scores and reasoning into a structured report.

Quick Start

Prerequisites

• Python 3.14+
• uv (Recommended for dependency management)

Installation

1. Set up the environment:

   # Create a virtual environment with Python 3.14
   uv venv --python 3.14
   source .venv/bin/activate

2. Install Vera Core:

   # Install the core engine from the local directory
   uv pip install "git+https://github.com/google/vera.git#subdirectory=vera"
   
   # Or if you already cloned the repo
   uv pip install vera

3. Install an Example Plugin: Vera requires plugins to define what to test. Let's install the SQL Query Assistant example.

   uv pip install "git+https://github.com/google/vera.git#subdirectory=plugin_example/vera_sql_query_assistant"
   
   # Or from the local folder
   uv pip install plugin_example/vera_sql_query_assistant

Running a Test

1. Configure API Key: Vera uses Gemini as the default judge. You'll need an API key.

   vera config -k
   # Follow the prompt to enter your key securely

2. Verify Setup: Check if the plugin is recognized.

   vera list
   # Should show: vera_sql_query_assistant

3. Run the Evaluation:

   # Run tests and save results to ./out
   vera test --dst-dir ./out

You can now open the generated CSV file in the out/ directory to analyze the results.

---

Maintained by the Vera Team.