I am very new to game development but I was able to scour the internet to figure out Perlin noise enough to implement a very simple 2D tile infinite procedural world.

Here's the question and it's more conceptual than code-based in answer, I think. I understand the concept of "I plug in (x, y) and get back from Perlin noise p" (I'll call it p). P will always be the same value for the same (x, y) (as long as the Perlin algorithm parameters haven't changed, like altering number of octaves, et cetera).

What I want to do is be able to zoom into a square and be able to generate smaller squares inside of the already generated overhead tile of terrain.

Let's say I have a jungle tile for overhead terrain but I want to zoom in and maybe see a small river tile that would only be a creek and not large enough to be a full "big tile" of water in the overhead. Of course, I want the same net effect as a Perlin equation inside a Perlin equation if that makes sense? (aka. I want two people playing the game with the same settings to get the same terrain and details every time).

I can conceptually wrap my head around the large tile being based on an "zoomed out" coordinate leaving enough room to drill into but this approach doesn't make sense in my head (maybe I'm wrong). I'm guessing with this approach my overhead terrain would lose all of the cohesiveness delivered by the Perlin. Imagine I calculate (0, 0) as overhead tile 1 and then to the east of that I plug in (50, 0). OK, great, I now have 49 pixels of detail I could then "drill down" into. The issue I have in my head with this approach (without attempting it) is that there's no guarantee from my Perlin noise that (0,0) would be a good neighbor to (50,0) as they could have wildly different "elevations" or p/resultant values returning from the Perlin equation when I generate the overhead map.

I think I can conceive of using the Perlin noise for the overhead tile to then reuse the p value as a seed for the "detail" level of noise once I zoom in. That would ensure my detail Perlin is always the same configuration for (0,0), (1,0), etc. ad nauseam but I'm not sure if there are better approaches out there or if this is a sound approach at all.