Welcome to this latest cohort, where we're going to be building a personal assistant in TypeScript. I wanted to build this course because we had such a good response to the previous cohort, but there was some stuff missing from there that I wanted to cover in depth.
Specifically, and this is the big one: retrieval. Being able to take an existing data set and plug it into an LLM, in other words, retrieve from that data set in response to questions, is one of the most powerful things you can do as an AI engineer.
You can build all sorts of cool things with retrieval:
A lot of people ask me: should I train my own LLM in order to specialize it? No, probably not. You just need to get really good at retrieving from a private data source.
We've got two full days on retrieval in this cohort.
On the third day, we're going to touch on evals. I know we're building personal software here, but really, every time you're building personal software, you should be training for production.
Building evals is a key part of putting AI apps into production and making them better once they're there. We're going to learn how to evaluate retrieval systems as well as more complex tool calling systems too.
On day four, we're going to touch on something that I've never touched on before: memory. The whole idea of a personal assistant is that it learns from your preferences, from your behaviors, and from what you actually do. In theory, it should get better over time.
So in day four, we're going to build an assistant that actually learns from you, and we're going to look at all sorts of cool things you can do with memory.
On the fifth day, we're going to focus on Human in the Loop, which is an incredibly important pattern for allowing your AI to actually do things in the world.
If you've taken the AI SDK v5 Crash Course, you know about tools, but how do you actually grant access to powerful tools to the LLM while keeping the user in full control?
The answer is Human in the Loop, and we're going to build it from scratch using the AI SDK's custom data parts API to the absolute maximum.
Once you finish that section, you'll know the custom data parts API back to front and you'll be able to build all sorts of powerful user interfaces with it.
Best of luck in the cohort. I'll see you in the office hours, and I'll be around in the Discord if you have any more questions. Nice work and I'll see you in the next one.