Before we get into the prompts it’s important to understand the Chat Completion API structures, specifically the messages array. 1st – I’d recommend that you take a close look at the Azure OpenAI REST API reference document. 2nd – I would recommend that you play with the Chat Completion endpoints using Postman. You can find a Postman Chat Completion Collection in my AI Fundamentals repo.
Basically, everything is a Plugin! Plugins are the building blocks of your AI solution. It allows you to define the tasks the Kernel should complete and allows you to augment the capabilities of the LLM and your solution. Plugins give you the ability to add native code. You can use built-in plugins or build your own, you can pass arguments and the functions of your plugins can be auto invoked.
There are multiple ways to work with prompts in Semantic Kernel. In this section we will dive into the various ways you can work with prompts. The most powerful way is to use the SK Prompt Templating Language, but you can also use Interpolated Strings. You can inject the Chat Completion Message structure and SK will automatically convert it into the proper Message Array. You can use inline prompts are load them from file. When using the SK Template Language you can invoke plugins directly from the prompt. You can pass arguments into the prompt and you can actually load a prompt as a Plugin and use Auto Invoke to call it as well.
If you want make your solution more robust, you can load your prompts from file. Loading prompts from file allows you to alter or change the prompts without having to redeploy the code base.
When loading prompts from file Semantic Kernel requires you to have two files 1) SKPrompt.txt 2) Config.json. You can take a look at the Lesson-2 solution for details on how to do this.
