That's why it's idiomatic in OCaml to write interface files which contain type annotations for the modules. In practice many or most modules (over time) are given interface files which strongly contain the area of effect of type inference.

Is it? I've only done a bit of OCaml (working on Sail) but that doesn't have any type interface files.

In virt-v2v we eventually enforced that every module file also has a corresponding interface file: https://github.com/libguestfs/virt-v2v/blob/master/check-mli...

Probably if you need to do that it should just be enforced by the compiler. It's pretty clear to me (and the author of this post) that Rust got this right - small scale type inference is fine (in function bodies) but you really want explicit types for functions and up.

Without that, you lose out massively on readability and error message quality.

In Sail, I have IDE type inlays working but it barely helps with reachability because half the types are inferred as generic.

As a user of OCaml I have to disagree. Rust requires far too many type annotations and they are mostly unnecessary and clutter everything up (particularly with it's very heavy syntax). In OCaml my editor shows me types if I mouse over a variable, and if I need to document the code I can do that, but I don't need to.

My point is it isn't very useful if half the types are 'a

We do have type interface files for almost every .ml file

I know this is an unpopular position in the OCaml ecosystem, but I find .mli files to be in all ways inferior to the "inline" approach Rust (amongst others) take. OCaml requires perpetual context switches between .ml and .mli files, which can be alleviated with proper IDE tools but is still annoying, but most annoyingly requires writing every type definition and module signature twice.

On the other hand the OCaml/interface files approach is superior when you want to read documentation w/o implementation, but in practice using odoc is superior because you get hyperlinks.

In 10 years of using OCaml, I think everytime I interacted with mli I'd rather the author annotated types at the definition site.

That's funny because I feel the opposite. An interface file saves me from having to wade through the implementation details and lets me focus on the exact parts that are relevant to me as a reader of the code. And yes odoc is nice too but the source of its usefulness usually comes from the fact that it's generated from a carefully curated interface file.

Would type "dehints" be the best of both worlds: you IDE could hide all the types (just like with type hints it adds them) and you could toggle them to eg hide when reading the code? But then also you could toggle them back when needed instead of having to open an interface file?

OCaml IDEs including emacs can do that already, and yes it's a good thing.

In my opinion the best of both worlds would just be to have Rust-style inline annotations and an "interface mode" that hides implementations (most modern IDEs can already do that) and non-public function.

BTW I am not sure why you would want to hide the types, I find them always useful.

I agree that the usefulness of documentation systems such as odoc comes from a carefully curated interface, but I disagree with them needing a separate file which in my experience is cause of friction.

Other languages (I am familiar with Rust and Python) also have documentation systems that only show you public methods (and functions, types, etc.), and I find them equally useful as odoc. They just take the visibility information from inline annotations (which are enforced by the compiler, in the case of Rust) instead of a separate file.

> I find .mli files to be in all ways inferior to the "inline" approach Rust (amongst others) take

Agreed. The same criticism can be levelled at C and C++ with header and implementation files.

I disagree with the article here (and it’s a good article overall!) The compiler simply needs to be armed to detect this pattern and give better error messages. Vast majority of cases the user simply forgot an argument.

* * *