LaTeX Intelligence

LaTeX Intelligence is a VS Code extension that does two things.

It turns a LaTeX paper into a Word document you can actually edit. Equations stay equations, cross-references stay live, tables keep their shape, and the page keeps the layout your document class gives it. It works with no TeX installation at all.

It tells you why a build failed. The exact line, what the error means in plain words, and what is missing from your machine — including the usual case where the same file compiles in a terminal but not in the editor.

Left: compiled by pdflatex. Right: the exported .docx, opened in Word.

Install

Search for LaTeX Intelligence in the VS Code Extensions view, or:

ext install MudassirUddinMohammed.latex-intelligence

Nothing else is required to export to Word. To build PDFs you need TeX Live or MiKTeX.

Getting started

1. Open a .tex file. The extension activates on its own.

2. Build the PDF. Click the ▶ button at the top right of the editor.

The PDF is written next to your .tex file. Build again without changing anything and it finishes in about a hundredth of a second, because nothing it read has changed.

3. Export to Word. Click the document button beside ▶.

The .docx is written next to your .tex file too. A report opens in the Output panel listing anything that did not convert cleanly — and it is specific: "figure 7 is a TikZ diagram and no LaTeX engine is installed, so it is a placeholder", not "the conversion may have issues".

4. If a build fails, press Ctrl+Shift+P and run Explain build errors.

5. Before you submit, click the LaTeX Intelligence icon in the activity bar on the left. Three panels: paper health, build profile, and a preflight check against your target venue.

Examples

A paper with equations and cross-references

\documentclass[conference]{IEEEtran}
\begin{document}
\section{Method}
\label{sec:method}

\begin{equation}
  \label{eq:drift}
  d_i = h^{\text{doc}}_i - h^{\text{pdf}}_i
\end{equation}

Section~\ref{sec:method} defines the drift in \eqref{eq:drift}.
\end{document}

In the exported document the equation is a real Word equation you can click into and edit. \ref and \eqref become live Word fields, so if a collaborator inserts a section above, the numbers renumber themselves.

A diagram, with no LaTeX installed

\begin{tikzpicture}
  \node[draw, rounded corners] (a) {Parse};
  \node[draw, rounded corners, right=of a] (b) {Render};
  \draw[->] (a) -- (b);
\end{tikzpicture}

With a LaTeX engine available the picture is compiled against your own preamble and embedded — pixel for pixel what the paper prints. With no engine it is interpreted and drawn as native Word shapes you can move and recolour.

A table

\begin{tabular}{lrr}
  \toprule
  Class & Paper & Ours \\
  \midrule
  IEEE conference & 143 & 126 \\
  \bottomrule
\end{tabular}

Each column comes out as wide as its own content, the way LaTeX sets one — not four equal columns with a wrapped heading. booktabs rule weights, \multicolumn spans, \multirow merges and cell shading all survive.

What survives the conversion

Equations become OMML, Word's own format. Editable, not pictures. Fractions, radicals, matrices, cases, n-ary operators, accents, delimiters.

Cross-references and citations become live Word fields, carrying the numbers your .aux file printed, so the document reads correctly before anyone updates a field.

Figures — TikZ, pgfplots, circuitikz, PDF, EPS and SVG. Vector wherever the tools allow it.

The page comes from your document class. IEEEtran, ACM, APA, MLA, Chicago, Vancouver, Springer LNCS and Elsevier each get their own margins, headings, caption labels and citation form, measured out of the class files rather than guessed.

The title block is always centred, whatever the class does.

Every file is checked against the schema Word enforces before it is written, so you never meet the "unreadable content" dialog.

What you need

Word export — nothing at all.

PDF builds — TeX Live or MiKTeX.

TikZ figures embedded as images — a LaTeX engine and a rasteriser: Poppler, Ghostscript or ImageMagick. Without them the diagrams are still drawn, as Word shapes.

SVG figures kept as vector — Inkscape.

Run Doctor: check toolchain to see what you have and what it enables.

Commands

Build PDF — compiles, picks the right bibliography program, and reruns until the references settle.

Export to editable DOCX — the main path.

Explain build errors — the real cause, in plain words, instead of the raw log.

Doctor: diagnose environment and PATH — for "it works in the terminal but not in VS Code". Compares the two PATHs and names the directory VS Code is missing.

Doctor: check toolchain — what is installed.

Run paper health check — missing citations, broken references, duplicate labels, uncaptioned floats.

Run journal preflight — against IEEE, ACM, NeurIPS, ICML, Springer, Elsevier or CVPR.

Also: Select main .tex file, Show build profile, Show last DOCX conversion report, Write paper.lock, Check Overleaf bridge integrity, Clear figure render cache.

Settings

Everything works without configuring anything. The ones worth knowing:

latexIntelligence.outputDirectory — where the PDF and DOCX go. Empty by default, which means beside your .tex file.

latexIntelligence.engine — auto, and it honours a %!TEX program line.

latexIntelligence.mainFile — auto-detected by looking for \documentclass.

latexIntelligence.health.runOnSave — on by default.

The rest are in the Settings UI under "LaTeX Intelligence".

Known limitations

The bounds on a \sum in a numbered equation sit beside the sign rather than above and below it. Word lays out inline maths that way, and the equation number has to share the paragraph.

Table captions are numbered 1, 2, 3 rather than I, II, III in classes that use Roman numerals.

Three of the eight papers in the internal test corpus run a page longer than the compiled PDF. The cause is body text density rather than any one element.

Contributing

The source is on GitHub: https://github.com/Mudassiruddin7/LaTex-Intelligence

To work on it:

git clone https://github.com/Mudassiruddin7/LaTex-Intelligence.git
cd LaTex-Intelligence
npm install
npx tsc --noEmit

Press F5 in VS Code to launch a second window with the extension loaded, and open a .tex file in it.

To send a change: fork the repository, make a branch, and open a pull request. Two things a reviewer will look for.

Check the output by looking at it. Word does not do what the markup says it will — it applies its own row heights, shrinks shapes to fit, and re-lays maths its own way. If you change how something is rendered, export a document, open it in Word, and say in the pull request what you saw.

Take numbers from a measurement. If you add a document class, read its page and its font sizes out of the class file rather than from a style guide.

Feedback

Something wrong, or missing? Tell me here — it goes to a form I read.

Bugs are also welcome as GitHub issues.

MIT licensed.