Is it possible to train GitHub Copilet for an entire codebase?

[Nefcanto]

Select Topic Area

Question

Body

Explain your requirements to GitHub Copilot, and you get a line of code, a function, or a class. That's awesome. But that's just the start of a long value-chain.

The true value comes when GitHub Copilot becomes the next team member.

The problem with AI code generation is that it's generic. It does not encompass the architecture, the coding style, the conventions, and many more aspects of a team. So a lot of times when we use them we have to adjust them for our code.

Our experience as a team is that AI code generation is like fast StackOverflow. On SO you had to wait to get answers, here you get them quicky. But from that point onwards the path is the same. Most of the time you can't just copy/paste the code directory into your code. You need to adjust:

1. Naming conventions including casing and semantics
2. Coding style (example => every attribute on a separate line in JSX)
3. Architecture (example => three-layered module in a modularized monolith without dependency injection, using these base classes)
4. Poicies (example => do not use fetch, use axios instead)

And if you get into prompts rephrasing to get the code nearer to your codebase style, most of the time the explanation time is not worth it. You can do it manually faster.

What we have understood is that the true value comes when the AI solves the problem in the context of our codebase.

Is that possible with GitHub Copilot? Can we feed it our entire codebase and ask it to solve problems according to the current codebase?