How to A/B test LLM models and prompts

Jul 01, 2024

Posted by

Lior Neu-ner

On this page

1. Download the sample app
2. Add PostHog to your app
3. Capture events for your A/B test
4. Create an A/B test in PostHog
5. Implement the A/B test code
Further reading

A/B tests enable you to compare how changes to LLM models and prompts affect your app. In this tutorial, we'll show you how to set one up to effectively evaluate your LLM improvements.

We set up a basic Next.js app, add PostHog, create A/B tests, and implement the code for them.

While this tutorial focuses on Next.js and Node, PostHog supports many different SDKs and frameworks. The concepts in this tutorial apply to all our supported SDKs and frameworks.

1. Download the sample app

We've created a basic recipe builder app for this tutorial. You can download it from Github.

Terminal

git clone https://github.com/PostHog/llm-ab-tests-sample-app
/```

To set your app up, first ensure [Node](https://nodejs.dev/en/learn/how-to-install-nodejs/) is install. Then run `npm install` to install all dependencies. 

You must also replace `YOUR_API_KEY` in `src/app/api/generate-recipe/route.js` with your Open AI api key:

```js file=src/app/api/generate-recipe/route.js
import OpenAI from 'openai';

const openai = new OpenAI({
  apiKey: "YOUR_API_KEY",
});

Run npm run dev and go to http://localhost:3000 to everything in action. The app works as follows:

Users first log in by entering in their email. The code for this page is in src/app/page.js.
After login, they are navigated to recipe-builder/page.js. Here they can type in a few ingredients and receive a recipe created by an LLM.
Once they've received the recipe, they can rate the response.
API calls to the LLM are made in src/app/api/generate-recipe/route.js. We're using ChatGPT for this tutorial, but you can use any LLM you like.

2. Add PostHog to your app

With our app set up, it’s time to install and set up PostHog. We install both the JavaScript Web and Node SDKs:

The JavaScript Web SDK runs client-side. We use this to capture user evaluations for the helpfulness of the recipes.
The Node SDK runs server-side. We use this to implement our A/B test code since our LLM requests run in the backend.

Run the following commands in your root directory:

Terminal

# in the root directory
npm install posthog-node
npm install posthog-js
touch src/app/providers.js

This installs the SDKs as well as creates a providers.js file. We use this to set up a PostHog provider for our client-side code. Add the following code to the file:

app/providers.js

'use client'
import posthog from 'posthog-js'
import { PostHogProvider } from 'posthog-js/react'
import { useEffect } from 'react'

export function PHProvider({ children }) {
  useEffect(() => {
    posthog.init('<ph_project_api_key>', {
      api_host: 'https://us.i.posthog.com',
      person_profiles: 'identified_only'
    })
  }, []);
  return <PostHogProvider client={posthog}>{children}</PostHogProvider>
}

Add your PostHog API key and host to the above code. You can find them in your project settings.

Next, we import the PHProvider component into our app/layout.js and wrap our app with it:

app/layout.js

import "./globals.css";
import { PHProvider } from './providers'

export default function RootLayout({ children }) {
  return (
    <html lang="en">
      <PHProvider>
        <body>{children}</body>
      </PHProvider>
    </html>
  );
}

To test everything has been set up correctly, restart your app and click on a few buttons. You should start seeing events in your PostHog activity tab.

Lastly, to set PostHog up in our server-side API route, import posthog-node and initialize it at the top of the src/app/api/generate-recipe/route.js file using your API key and host:

src/app/api/generate-recipe/route.js

import OpenAI from 'openai';
import { PostHog } from 'posthog-node'

const openai = new OpenAI({
  apiKey: "YOUR_API_KEY",
});

export async function POST(request) {
  const posthog = new PostHog("<ph_project_api_key>", {
    api_host:  "https://us.i.posthog.com",
  })
  
// rest of your existing code

3. Capture events for your A/B test

Now that we have PostHog set up, the next step is to capture events to use as our goal metric in our A/B test. In our case, we want to use a user feedback score as the goal metric.

To do this, we capture an event every time clicks the Yes or No buttons when asked if the LLM response was helpful. We assign a score of +1 if they click Yes and -1 if No.

Update the code in recipe-builder/page.js to import the usePostHog() hook and capture events in the handleYesClick and handleNoClick functions:

recipe-builder/page.js

'use client';

import { useState, useEffect } from 'react';
import { useRouter } from 'next/navigation';
import { usePostHog } from 'posthog-js/react';

export default function RecipeBuilder() {
    // existing code...

  const posthog = usePostHog()

  const handleYesClick = () => {
    posthog.capture('user_recipe_feedback',
      {
        score: 1
      }
    )
  };

  const handleNoClick = () => {
    posthog.capture('user_recipe_feedback',
      {
        score: -1
      }
    )
  };

  // rest of your existing code

Next, we want to make sure we attribute these events to the user who submitted them. This ensures our A/B test results are accurate.

To do this, we call posthog.identify() with the user's email. This links any captured events to the user.

Update the code src/app/page.js to import the usePostHog hook and call identify inside the handleLogin function:

src/app/page.js

'use client';

import { useEffect, useState } from 'react';
import { useRouter } from 'next/navigation';
import { usePostHog } from 'posthog-js/react';

export default function Home() {
  // existing code

  const posthog = usePostHog()

  const handleLogin = () => {
    if (email) {
      posthog.identify(email);
      localStorage.setItem('email', email);
      router.push('/recipe-builder');
    }
  };

  // rest of your existing code

Lastly, when a user log outs we call posthog.reset() to unlink any future events. Update the handleLogout function in recipe-builder/page.js to do this:

recipe-builder/page.js

// existing code...
  
  const handleLogout = () => {
    posthog.reset();
    localStorage.removeItem('email');
    router.push('/');
  };

  // existing code...

Restart your app and test everything is working by clicking the Yes and No buttons a few times in your app. You should see user_recipe_feedback events with the user's email in your activity tab.

4. Create an A/B test in PostHog

We're going to create a multivariate experiment that tests both the model and prompt. It will have the following variants:

Control: uses the model gpt-3.5-turbo and the prompt Generate a recipe using these ingredients.
Model change: Same prompt as control but uses the gpt-4o model.
Prompt change: Same model as control but uses the prompt Generate a healthy and delicious high protein recipe using these ingredients

To do this, go to the A/B testing tab and create an A/B test by clicking the New experiment button. Add the following details to your experiment:

Name it "LLM experiment".
Set "Feature flag key" to llm-experiment.
Click + Add test variant. Name the first test variant to model-change and the second one prompt-change.
Click Continue.
Change the goal type to Trend.
Under goal criteria, select the user_recipe_feedback event. Click on the total count dropdown and select Property value - average. Select score as the property.

Click Save as draft and then Launch

5. Implement the A/B test code

To implement the A/B test, we:

Fetch the llm-experiment flag in api/generate-recipe/route.js using posthog.getFeatureFlag()
Update the model or prompt depending on the flag value.
Call await posthog.shutdown() at the end of our request to flush and send all pending events.

api/generate-recipe/route.js

// your existing imports and code

export async function POST(request) {
  const posthog = new PostHog("<ph_project_api_key>", {
    api_host:  "https://us.i.posthog.com",
  })
  
  const { ingredients, email } = await request.json();

  try {
    let model = "gpt-3.5-turbo";
    let prompt = "Generate a recipe using these ingredients";

    // Fetch the experiment variant and update the model or prompt if needed
    const enabledVariant = await posthog.getFeatureFlag('llm-experiment', email);
    if (enabledVariant === 'model-change') {
      model = 'gpt-4o';
    } else if (enabledVariant === 'prompt-change') {
      prompt = 'Generate a healthy and delicious high protein recipe using these ingredients';
    }

    const completion = await openai.chat.completions.create({
      model,
      messages: [
        { role: "system", content: "You are a helpful assistant that generates recipes." },
        { role: "user", content: `${prompt}: ${ingredients}` }
      ],
    });

    // Flush and send all pending events before the end of the request
    await posthog.shutdown()

    return new Response(JSON.stringify({ recipe: completion.choices[0].message.content }), {
      headers: { 'Content-Type': 'application/json' },
    });
  } catch (error) {
    console.error('OpenAI API error:', error);
    return new Response(JSON.stringify({ error: 'Failed to generate recipe' }), {
      status: 500,
      headers: { 'Content-Type': 'application/json' },
    });
  }
}

And we're done setting up our A/B test! Open your app, log in with a few different emails, and click the Yes and No buttons a few times. You should now see results in your experiment page.

How to A/B test LLM models and prompts

1. Download the sample app

2. Add PostHog to your app

3. Capture events for your A/B test

4. Create an A/B test in PostHog

5. Implement the A/B test code

Further reading

Product for Engineers

Comments