Getting started with Braintrust

Install Braintrust libraries

First, install the Braintrust SDK (Typescript or Python).

npm install braintrust autoevals

yarn add braintrust autoevals

Create a simple evaluation script

Next, create a file named tutorial.eval.ts or eval_tutorial.py with the following code.

import { Eval } from "braintrust";
import { LevenshteinScorer } from "autoevals";
 
Eval(
  "Say Hi Bot", // Replace with your project name
  {
    data: () => {
      return [
        {
          input: "Foo",
          expected: "Hi Foo",
        },
        {
          input: "Bar",
          expected: "Hello Bar",
        },
      ]; // Replace with your eval dataset
    },
    task: async (input) => {
      return "Hi " + input; // Replace with your LLM call
    },
    scores: [LevenshteinScorer],
  },
);

This script sets up the basic scaffolding of an evaluation:

• data is an array or iterator of data you'll evaluate
• task is a function that takes in an input and returns an output
• scores is an array of scoring functions that will be used to score the tasks's output

(You can also write your own code. Make sure to follow the naming conventions for your language. Typescript files should be named *.eval.ts and Python files should be named eval_*.py.)

Create an API key

Next, create an API key to authenticate your evaluation script. You can create an API key in the settings page.

Run your evaluation script

Run your evaluation script with the following command:

BRAINTRUST_API_KEY=<YOUR_API_KEY> npx braintrust eval tutorial.eval.ts

This will create an experiment in Braintrust. Once the command runs, you'll see a link to your experiment.

View your results

Congrats, you just ran an eval! You should see a dashboard like this when you load your experiment. This view is called the experiment view, and as you use Braintrust, we hope it becomes your trusty companion each time you change your code and want to run an eval.

The experiment view allows you to look at high level metrics for performance, dig into individual examples, and compare your LLM app's performance over time.

Run another experiment

Now that you've run your first eval, you'll notice that we only achieved a 77.8% score. Can you improve that?