iTranslated by AI
Book Review: Introduction to Building ChatGPT/LLM Systems with Azure OpenAI Service
I recently received a complimentary copy of "Azure OpenAI Service Introduction to ChatGPT/LLM System Construction" (scheduled for release on January 24, 2024) from the author. This article serves as a review of that book.
Note: Although many of the authors are employees of Microsoft Japan, the writing is an individual activity of the authors and does not appear to be an official Microsoft Japan publication.
General Review
This is a book where you can systematically learn about Azure OpenAI. It is a must-read for anyone planning to develop applications using Azure OpenAI.
Furthermore, it doesn't stop at just explaining Azure OpenAI itself; many pages are dedicated to integration with the Azure ecosystem—such as private networking using VNet and load balancing using Azure Load Balancer—which are often highlighted as benefits of using Azure OpenAI. It was impressive that this book focuses not just on Azure OpenAI as a technical component, but on building actual solutions using Azure OpenAI.
Many companies are implementing internal enterprise searches using Azure AI Search, and I thought it was excellent that a whole chapter is dedicated to RAG, the technical method used for this, covering everything from implementation to evaluation. In particular, evaluation methods for RAG are still evolving, and there are few books that cover the technical foundations, making this a must-read for those looking to implement RAG-based search.
Additionally, for those wondering, "I hear a lot about Copilot, but can we build one ourselves?", understanding the contents of this book will help you grasp how Copilot works.
Who I Recommend This Book To
- Those who want to learn Azure OpenAI systematically
- Those who want to know not only about Azure OpenAI but also about the Azure services frequently used alongside it
- Those who are planning to build solutions that incorporate Azure OpenAI
People for whom this book might be a mismatch
Every book has an intended audience. The following is not intended as a criticism of the book's content, but rather to point out that you might fall outside the intended audience if you read it with certain expectations. I will also include my personal views on what kind of learning materials might be better suited for those cases.
-
Those who want to know about DALL-E or Whisper within OpenAI
This book provides a detailed explanation of GPT-3.5 and GPT-4. Therefore, it might be a mismatch for those who want to learn specifically about DALL-E or Whisper. That said, since prompt engineering is necessary for both, reading this book before moving on to DALL-E or Whisper is a viable path. However, since there are currently no books for systematically learning DALL-E or Whisper, I recommend reading the official documentation. -
Those who want to know about the latest prompt engineering
While well-established prompt engineering techniques are summarized nicely, the latest methods (such as the well-known "Emotion Prompt") are not covered. Consequently, those already familiar with the field might find much of the content familiar. Since it is often difficult to find the latest information in printed books, I recommend looking through arXiv or following the latest updates on X (Twitter). -
Those who want to know about Content Filters in detail
The Content Filter is a powerful, Azure OpenAI-specific feature. While there is an explanation of what it is, the book only briefly touches upon how to integrate it into applications or how to handle scenarios when a filter is triggered. Those specifically looking for deep dives into this area might find it insufficient.https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/content-filter
-
Those who want to know about adversarial prompts
Related to content filters, the book does not cover adversarial prompts—attack methods against LLM applications such as prompt injection or prompt leaking—or how to defend against them. While turning on Jailbreak detection in the Content Filter is a start, the book does not discuss whether these can be prevented through service design or specific prompting techniques. -
Those who want to learn LangChain or Semantic Kernel systematically
While this book mentions LangChain and Semantic Kernel, it does so only as an introduction and is not structured for systematic learning of these frameworks. If you want to learn them systematically, you should look for dedicated books. However, these libraries are updated very frequently. Although stable versions were recently released, I personally feel it's debatable whether books (which take time to publish) are the best way to learn them. Starting with the official documentation's quickstart is likely the best first step. -
Those who want to do intensive hands-on training
This book provides small sample code for each feature, but it is not structured around building a single complete application to learn Azure OpenAI. Therefore, those looking for an intensive hands-on experience might find it lacking. However, I personally prefer the "learn the basics and imagine the applications" approach, so I find the book's structure quite appealing.
Particularly Good Points
-
Covers ethics beyond technology
It describes how to address the ethical challenges of LLMs based on Microsoft's "Responsible AI" framework. Since technical books often focus solely on technical elements, I found this inclusion very valuable. -
Covers evaluation methods
While many organizations are currently in the "let's just try building it" phase, as part of corporate activities, the time will eventually (or already has) come to consider "how to evaluate this LLM application." Personally, I've been focusing on this area recently, so I found it very good that even evaluation methods were covered.
Side Notes
The following are purely personal impressions.
-
The appendix on how ChatGPT works is interesting
My knowledge of machine learning had stopped around the time of LSTM when I obtained the JDLA E Certificate, so it was very helpful that the delta from that point was included as an appendix. -
Makes me crave French crullers
Is that why there's a donut on the cover?
Discussion