Show HN: Tylax – A bidirectional LaTeX to Typst converter in Rust

27 pointsposted a month ago
by democat
(github.com)

9 Comments

holg

a month ago

nice job! so i tried around and so far am impressed, anyhow:

\newcommand{\foo}[1]{\bar{#1}} \renewcommand{\bar}[1]{\foo{#1}} % mutual recursion \foo{x} \def\x{\y}\def\y{z}\x % chained expansion

+>

#set page(paper: "a4") #set heading(numbering: "1.") #set math.equation(numbering: "(1)") /* \foo / / \y /

\foo{x} → \bar{x} → \foo{x} → ∞ Expected: hit depth limit, emit warning, output either x or the unexpanded \foo{x} Actual: / \foo */ — silently converted to a comment, lost the argument x entirely

democat

a month ago

Thanks for the stress test! You are absolutely right, and I really appreciate you catching this edge case.

I've traced the issue in the codebase: 1. The recursion depth limit *is* triggering correctly (preventing an infinite loop). 2. However, when it bails out, it returns the partially expanded macro call (e.g., `\foo{x}`). 3. Since the macro definition was removed during the preprocessing step, the parser then sees `\foo` as an unknown command and converts it into an error comment, accidentally discarding the argument `{x}` in the process.

*The intended behavior* should definitely be to preserve the content (the `x`) even if the macro logic fails. I will fix the fallback logic to ensure it fails gracefully without eating the arguments.

Thanks again for the sharp eye! These kinds of checks are super helpful.

leephillips

a month ago

Could you briefly explain why Pandoc was not sufficient? (Obviously it can’t do the TikZ -> CeTZ conversion.)

democat

a month ago

Great question! I love Pandoc and use it often, but as a "universal" converter, it sometimes misses the nuances of specific pairs. Tylax is designed specifically for the LaTeX $\leftrightarrow$ Typst workflow. By focusing on just this 1-on-1 pair, we can offer: Better Math & Macros: A built-in macro expander handles custom commands (\newcommand) and complex nested math that general parsers often struggle with. Cleaner Code: The output is designed to be idiomatic and human-readable (e.g., using native Typst functions), not just "compilable." WASM Support: Being written in Rust means it runs instantly in the browser, making it easy to embed in web apps without a backend. Pandoc is the Swiss Army knife; we're trying to be a specialized tool just for this specific transition.

fiddlosopher

a month ago

Pandoc does know how to expand LaTeX macros. For example, given the LaTeX

  \newcommand{\pair}[2]{\langle #1, #2\rangle}
  $$\pair{a^2}{\frac{\pi}{2}}$$
pandoc will give you the Typst

  $ chevron.l a^2 \, pi / 2 chevron.r $
which is correct. Tylax, on the other hand, seems to have problems with this example, producing

  $ angle.l^()frac(pi,)angle.r  $
which does not compile with typst. Going the other direction, pandoc also understands typst scripting. For example, from

  #let count = 8
  #let nums = range(1, count + 1)
  #let fib(n) = (
    if n <= 2 { 1 }
    else { fib(n - 1) + fib(n - 2) }
  )

  The first #count numbers of the sequence are:

  #align(center, table(
    columns: count,
    ..nums.map(n => $F_#n$),
    ..nums.map(n => str(fib(n))),
  ))
pandoc produces this LaTeX:

  The first 8 numbers of the sequence are:

  {\def\LTcaptype{none} % do not increment counter
  \begin{longtable}[]{@{}llllllll@{}}
  \toprule\noalign{}
  \endhead
  \bottomrule\noalign{}
  \endlastfoot
  \(F_{1}\) & \(F_{2}\) & \(F_{3}\) & \(F_{4}\) & \(F_{5}\) & \(F_{6}\) &
  \(F_{7}\) & \(F_{8}\) \\
  1 & 1 & 2 & 3 & 5 & 8 & 13 & 21 \\
  \end{longtable}
  }
With the same input, Tylax produces:

  The first 8 numbers of the sequence are:

  \begin{center}

  \begin{tabular}{|c|}
  \hline
  \hline
  \end{tabular}\end{center}
which is just an empty table.

democat

a month ago

You are absolutely right. Thank you for pointing this out! Regarding your questions: 1. The `\pair` issue: This is definitely a bug. My macro expander is based on text replacement and obviously cannot handle nested parameters. I will fix the recursive logic. 2. The `fib` loop: Pandoc seems to use `typst-hs`, which contains a complete Typst evaluator. Tylax is strictly designed as a static AST transformer. We haven't implemented the Typst virtual machine, so loops or recursive functions cannot be executed. This will be gradually improved later to make it more usable; my claim of "better macro support" was clearly premature. This was a big mistake on my part, and we will strive to achieve this goal in future updates! Thank you very much for your feedback and for pointing out the bug!

leephillips

a month ago

That makes sense, especially the issue with macros. As many people have pointed out, since TeX is not just markup but an actual programming language, its output can not be determined, in the general case, without running the source through the TeX interpreter. Of course, the same is true of Typst.

democat

a month ago

You are absolutely spot on. Both systems are Turing-complete, so a perfect conversion without a full runtime execution is theoretically impossible for the general case. That's exactly the trade-off Tylax makes: we aren't trying to be a full TeX engine (which would be overkill and slow). Instead, we aim to cover the "99% use case" of academic and technical writing—where macros are mostly used for shorthand, notation aliases, or simple formatting, rather than complex computation. Our "limited macro expander" is the middle ground: it's dumb enough to be fast and safe (no infinite loops), but smart enough to handle the \newcommand shortcuts that riddle almost every paper. It's about being pragmatically useful rather than theoretically perfect

bayesnet

a month ago

I haven’t tried this out yet but my gripe with pandoc is that it produces latex (and typst) that no human would ever write. It looks messy and is annoying to share with coauthors.

This is not to say that pandoc isn’t a fantastic tool.