In words
What it is, why it matters, and what it is like.
Why am I learning this?
You are learning this because it is the first tool you will use to actually test an AI system you have built. Without it, you can only guess whether your changes helped or broke things. By the end of this page you will be able to: write a small evaluation file that runs a list of test prompts through a model and checks the answers; run that file from your terminal and see a pass/fail report; and set it up so that every time you change your code, the evaluation runs automatically. This is the same skill that professional teams use to stop a new model update from silently making their chatbot worse.
The idea, in plain terms
When you build an AI system, you want to know if it is getting better. You cannot just ask 'is it good?' because 'good' is too vague. Instead you write down a list of questions and the answers you expect, then you run your system on that list every time you make a change. If a change makes it answer a question wrong that it used to get right, you know instantly you broke something. This is what an evaluation harness does for you. Promptfoo and LightEval are two open-source tools that do this. They are not the only ones, but they are popular and they show the core idea: you write a config file that says what model to call, what prompts to send, and how to check the responses. The tool runs all of that for you and prints a report. The magic is that the config file is text, so it can be saved, shared, and run again later. That is what makes your testing reproducible: anyone can run the same file and get the same test.
An analogy
Imagine you are a chef creating a new recipe. You have a taste-tester (your AI model). You do not just ask the tester 'is it good?' because that is subjective and varies day to day. Instead you write a fixed tasting card: a list of dishes with the expected notes (sweet, salty, spicy). Each time you change the recipe, you serve the same tasting card and mark pass or fail for each dish. This is exactly what an evaluation harness does. The config file is the tasting card. The model is the taste-tester. The checks are the marks. Now, the analogy breaks down: a recipe change might improve some dishes and ruin others, and the tasting card captures that; but here you have hundreds of prompts, not ten dishes, and the checks are often automated, not human. Also, the model's answers are not just taste notes but free text, so the checks must be smart enough to tell if 'the sauce is too salty' means 'fail' or 'pass'.
Definition
Promptfoo and LightEval are open-source testing tools that let you define a list of prompts and expected outcomes in a configuration file, run them against a model, and automatically decide whether each response passes or fails, so you can catch regressions and compare versions.
Where this sits
You have not studied any other topics yet. This is your entry point into AI evaluation. After this, you will learn about Evaluation Datasets (the prompts and expectations themselves), Deterministic Validators (the simple checks that give pass/fail), and then LLM-as-Judge (when you use a model to grade another model). Promptfoo and LightEval are the harnesses that bring those ideas together into something you run.
