GSoC 2017: Documentation Browser for Juno

The aim of this GSoC project is to provide a convenient way to access documentation in the Juno IDE. Any work on this has to be on the Julia side (for getting the necessary information by introspection) and on the Atom side (for presenting said information).

Most of the work on the Julia side went into a new package, DocSeeker.jl, which implements all of the introspection necessary to get docstrings from installed packages; a small shim in Atom.jl (Atom.jl#99) then delegates any front end requests to that package.

The front end work is directly included in the Juno stack (atom-julia-client#377 and atom-ink#148, to be precise), since it’s pretty fundamental IDE functionality.

Backend: DocSeeker.jl

The two main challenges here are collecting docstrings and filtering docstrings, both in the most performant and reliable manner possible.

Collecting Information

DocSeeker.jl contains a function alldocs(), which will return information about all symbols available in the current Julia session. Those symbols are easily found by recursing through all currently loaded Modules and calling Base.names() on them. Additionally, Julia’s docsystem collects all symbols with attached docstrings, which can be easily retrieved.

All of that is pretty slow – it takes on the order of half a second on my machine with a couple of loaded packages (and returns information about ~13,000 symbols). At the same time the available symbols don’t change too often, so caching is a viable solution.

Filtering

There are all sorts of possible options to consider when filtering and searching through the symbols (and attached docstrings) returned by alldocs, but I’ve decided on a few that turned out to be most important while testing:

That last point warrants some more information, because it’s not as trivial as the other two to get at least somewhat right. The basic problem here is a (fuzzy) full text search, which is what each search engine out there in the depth of the internet tries to do. Naturally there are quite a few (open source) implementations out there already: solr, lunr (which is used by docs.julialang.org), but also e.g. the FTS extension for SQL and many more.

In the early days of the summer I was trying a couple of things to get this to work properly, but shelling out to Java or JavaScript seemed like overkill for the problem at hand, while SQL.jl was giving me quite a hard time and doesn’t ship the FTS extension by default.

A custom implementation did not seem too hard at first, but requires a good scoring function that, given a search query needle, maps a docstring to a number between 0 and 1: At first I tried rolling my own string comparison function (with mixed success), but then I stumbled upon the excellent StringDistances.jl which does pretty much all I needed.

The scoring function is applied to all relevant symbols in a threaded loop (which gives a free 1.5x speedup on my machine); afterwards all applicable filters are applied and the top 20 results are returned.

Filtering and searching takes about 0.1s on my machine, which means that it’s almost negligible compared to the time necessary for retrieving the docstrings.

Frontend

Now that DocSeeker.jl has found the results we asked for, it’s time to display them in an appealing manner: search

If you’ve used Juno before you may notice the much improved markdown rendering (which is of course available all throughout Juno): There’s syntax highlighting, LaTeX rendering and lots of general improvements.

Apart from that the docpane UI shows most relevant information (type of the binding, defining module, whether the binding is exported etc.); a click on the binding will take you to the defining location and a click on the module will give some information on that. Links also generally work fine (external ones will open in your default browser, while those defined with Documenter.jl’s [link](@ref) syntax will start a new search).

Try it!

These features will have been integrated into Juno at the beginning of September 2017, so feel free to try them!

Acknowledgements

I’d like to thank Mike Innes for all the fruitful discussions about implementation and functionality, as well as his guidance on Julia/Juno development in general (well before GSoC even started).

Donate Now