"Not so important functionalities" There is absolutely no way to know that from the code.
What makes a functionality important has everything to do with how a user feels about it. The code doesn't know. There is code in your car right now. Some makes the radio work. Some makes the breaks work. What about either set of code tells an AI which will keep you alive? The comments? Who trusts those?
We should not forget that AI is really good at sounding like it knows what it's talking about. Not so good at actually knowing. AI decides things with random. It tries to make it sound right. People warn that sometimes AI hallucinates. Sometimes the coin flip is correct. Sometimes it's not. But the BS always sounds good.
Even a stopped clock is right twice a day. Is it any less a hallucination when it happens to be right?
What we have is code that implements features. It would be nice if we had a tool that could recognize where a features code has been scattered to. Color code by the feature it supports you'd make me a happy guy.
I see two avenues to pursue here. First, given the AI has the ability to know if the program will compile or not, what combinations of parts of the commit decompose into code that even compiles? That's a lot of computation but there's a chance it might make a for a coherent report about a large commit.
Second, does the code have tests? Tests are not exactly features but they're a close second. Mapping code to the tests they support might be an interesting report. Depends on the tests.
Now I call those two different avenues, but understand, being able to compile is itself a test. The art here is going to be avoiding over decomposing. You don't want every semicolon becoming a sub commit. We're trying to simulate a coder who commits every time it compiles and the tests pass (which is a good habit kids).
That's doable. But an AI that knows which features I care about is either getting lucky or reading more than the code.