So I have worked with OpenAI models in Azure and wanted to gain a broader insight into the different Azure AI Services. The AI-102 exam looked like a good fit because it covers a comprehensive range of services and skills in depth. I started in September, and from the start, it took around 4 months. I wasn’t in a hurry and worked on it for a few hours now and then, balancing it with other commitments.
The Online Course with Exercises
The base for my study was the free online course AI-102t00 from Microsoft that is targeted toward the AI-102 certification. It goes through the different AI Services and how you use them, including coding labs.
Microsoft provides code projects for use in visual studio code that you then build on in the labs against resources in azure that you create as part of the labs.
You can do the coding part in a virtual environment provided by Microsoft but you will save time if you have a local installation of VS code and work there.
I have 160$ of Azure credits a month to use, because I have MSDN ultimate from work. But most AI single services have a free tier in Azure and you can remove the resources after usage in Labs, so the cost in Azure will be minimal. You probably get credits when registering that will last the whole course.
The exercises primarily use the SDKs and the code is in .net 7. So it can be outdated sometimes compared to the newer SDK versions. Most are basically the same but OpenAI SDK has been rewritten. So use the exact version mentioned in the labs to start with.
I did the course one time and then went back to the course material a few times more.
Study Guide
The study guide specifies what areas the certification covers and how they are weighted. With Natural language processing as the biggest part.
Plan and manage an Azure AI solution (15–20%)
Implement content moderation solutions (10–15%)
Implement computer vision solutions (15–20%)
Implement natural language processing solutions (30–35%)
Implement knowledge mining and document intelligence solutions (10–15%)
Implement generative AI solutions (10–15%)
REST Interface
The AI services are called through REST. I used Postman to call services. It is a good idea to look at the URL because there will be questions that involve its structure. The services have different URLs structure and it also varies if it is a multi or single service, a subdomain or a regional endpoint. Many services can use both. Then there are the URLs for downloading Containers.
The rest interface also includes what headers are set for authentication as well as body structure for the call and body structure for the response.
Services also differ in if they return the result directly or through a result URL that should be called after the first call.
When you create a new collection in Postman, search and use the Azure Infrastructure template to import some of the services into postman.
Training Videos
Microsoft has a series in 13 parts for this certification.
Study Cram on youtube by John Savill. Watched it early on to get an overview.
Microsoft Exam Prep video series focuses on what areas the exam covers without going into detail.
Test Exams
Microsoft has a test exam on the certification page.
I have done a few different test exams and Measureup has the best quality. Unfortunately also the highest price, but search for promo codes. I got 30 days for half the price. There were 119 questions in total. They have a free test exam also.
The Exam
My exam started with a case study with 6 questions. Then mixed questions. I had a total of 56 questions. I noticed 40 minutes in that I was a bit behind and picked up the pace. I ended up with 15 minutes left. Since the exam is open book, you can use Microsoft Learn in the exam interface. But if you do it too early on, it might eat up your time to answer all questions. I corrected at least one question. A few others I checked had the right answer already. I ended up with 885 points out of 1000, but it did not feel easy and I used the full time.
Do a system check directly when you have booked the exam. My work laptop had services that could not be shut down and I had to use my private laptop.
Azure AI Evolves
Microsoft is constantly evolving its Azure AI offering. What previously was called Cognitive Search is now Azure AI Search. LUIS is replaced by Conversational Language Understanding. Forms recognizer is renamed document intelligence. Qna Maker is replaced by custom question and answering and so on. The market is moving fast and Azure AI with it.
Thus using old study material, online tests and Gen AI LLMs that have their data cap date october 2023, like OpenAI 4o in Azure can be confusing. So try to use updated information.
This evolution will probably be reflected in the yearly renewal, which is a good thing.
Starting Point for Business Application
One lab in particular stood out as a good starting point for business application, the one about using your own data with OpenAI. It goes through the process of creating a vector search index on company documents in AI Search and then using it with an OpenAI model for a RAG solution. A basic version is up and running in 30-40 minutes. The prompt is matched against the vector index. Documents about the identified topic are sent as context with the prompt to the OpenAI model that processes the prompt with the provided data.