How to test AI using AI

Leveraging OSS for Testing AI Based Applications

Home

www.robin-gupta.com

SLIDESMANIA.COM

About me

Home

Recent Tabs

About Me

www.robin-gupta.com

​

www.robin-gupta.com

Explorer, open source contributor, maker, author, international speaker and Dad at home.

More boring stuff about me here:

SLIDESMANIA.COM

Agenda...

Recent Tabs

www.robin-gupta.com

​

Home

Introduction

🙏

Understand the Why

🤔

When and How of Evals

🧮

Application under Test

💻

Rule based Evals

📐

Using AI for Evals

🤖

AI based reasoning

🤖🧠

CICT pipeline for AI apps

💀

QnA and Chocolates�🍫

Icebreaker

❄️

SLIDESMANIA.COM

But first , what is AI?

Home

Recent Tabs

Problem domain

www.robin-gupta.com

​

“Artificial intelligence (AI) refers to computer systems capable of performing complex tasks that historically only a human could do, such as reasoning, making decisions, or solving problems. “ - Coursera.

GenAI

Our focus today

SLIDESMANIA.COM

Our case study

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

• We built a RAG application which serves as a customer facing chat bot.
• Testing took longer than development, for checking:
• Context adherence
• Accuracy
• Bias
• Live in production and serving 1000+ answers so far
• Automation testing is as important as manual testing for AI

SLIDESMANIA.COM

Testing which is easier?

Home

Recent Tabs

Problem domain

www.robin-gupta.com

​

A

B

SLIDESMANIA.COM

Hallucinations and human-in-the-loop

Home

Recent Tabs

Problem domain

www.robin-gupta.com

​

Note: Observed on 25 Feb 2024

SLIDESMANIA.COM

Changing needs for software testing

Home

Recent Tabs

Problem domain

www.robin-gupta.com

​

SLIDESMANIA.COM

LLM Testing Eval

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

We can evaluate LLM outputs on:

• Performance (speed and functionality)

• Effectiveness (accuracy and utility)

• Quality (user experience and reliability)

SLIDESMANIA.COM

Automated Evals: What and When

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

What should you evaluate?

• Context adherence
• Context relevance
• Correctness
• Bias and Toxicity

​

When should you evaluate?

• After every change (bug fixes, feature updates, data changes)
• Pre-deployment (merges to prod branch, end of sprint, prior to shipping hotfix)
• Post-deployment (on demand business needs)

SLIDESMANIA.COM

Demo Application under test: Quiz generator

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

​

​

​

​

Quiz generator

LLM

Data

User

Input

Output

Input: Write a quiz about science

Output: Sure ! Here are three science questions for you:

1. True or False: Water slows down the speed of light.
2. What did Marie and Pierre Curie discover in Paris?
3. Where were the first refracting telescopes invented?

​

SLIDESMANIA.COM

Test Strategy for AI systems

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

Start

Quiz Generator

Quiz Bank

Prompt Template

User �Input

Contains words

Fail

Contains formatting

Fail

Check�Hallucination

Fail

Stop

No

No

No

Yes

Yes

Yes

1

2

3

Rule based Evals

AI based automated Evals

SLIDESMANIA.COM

CI/CT pipeline for AI apps

Home

Recent Tabs

Solution domain

​

www.robin-gupta.com

​

Change

Per-Commit evals

Merge to main

Pre-release evals

Deploy

Report

Dev/feature branch

Main/Prod branch

Human review

Human review

SLIDESMANIA.COM

Key takeaways-1

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

Problem

Solution

Tip

• Evaluating LLM based apps is hard as they are non-deterministic
• LLM hallucinations need to be checked for every feature

• We can use rule based evals for string/pattern matching
• We can use LLMs to test non-deterministic nature and hallucinations in LLM based apps.

Ensure that all eval results are passed through a human in the loop for review.

SLIDESMANIA.COM

Key takeaways-2

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

Basics

Advanced

God mode

Colab notebooks presented in this session:

• Automated Evals_basics.ipynb
• Automated Evals_advanced.ipynb
• Automated Evals_AdvancedReasoning.ipynb

• Langchain
• promptfoo
• ​

• GitHub - openai/evals: Evals is a framework for evaluating LLMs and LLM systems, and an open-source registry of benchmarks.
• [2307.03109] A Survey on Evaluation of Large Language Models

SLIDESMANIA.COM

QnA and prizes

Home

Recent Tabs

Solution domain

www.robin-gupta.com

​

SLIDESMANIA.COM