Page Canary Design Doc

This document covers the engineering design of Page Canary

• We will the gpt 3.5 LLM model for cost reasons

Bot algorithm

Loop

The core bot loop can be classified into the following category:

State detection

• Bot reads the state of the world
  • Native apis
  • Bot
    • OCR
    • Computer vision

Planning

Bot figures out what it should do:

Execution

• Bot executes on it's plan

Page Canary Scaling

Free trial

Free trial hosts run on the main host

Overview

• Each paid instance is it's own VM
• When a premium user signs up we provision a VM
• Users are given a dashboard to control the VM
• Users can spin down the VM

Pricing

The pricing from AWS is $6 / VM

• If the users billing fails, we spin down their host VM and send them an email

Provisioning

• Call VM host api
• Store in database
• Connect to user owner

To provision a new VM we need to call the API to create a new $6 VM for the user

Installation

• SSH to VM
• Download installation script
• Run installation script

VM selection

It was determined that the t4g.small instance

• 2 GB RAM

AI

How do we use artificial intelligence at Page Canary?

Large language models (LLMs)

• Issue prioritization
• Page content issue detection
• Duplicate issue detection
• Dynamic QA test plans

Computer vision

Object detection

• Open CV object detection is used

Recommended reading resources

• Observations running more than 5 million headless sessions a week (opens in a new tab)
• LLM Powered Autonomous Agents (opens in a new tab)

Helpful shell scripts

Determine jpeg quality

identify -format '%Q' path-to-image.jpg

Page Canary Lang Chain

https://www.langchain.com/ (opens in a new tab) https://js.langchain.com/docs/modules/agents/ (opens in a new tab) https://minimaxir.com/2023/07/langchain-problem/ (opens in a new tab)