Summary

Pandoc provides an interface for users to write programs (known as filters) which act on pandoc’s AST.

Pandoc consists of a set of readers and writers. When converting a document from one format to another, text is parsed by a reader into pandoc’s intermediate representation of the document—an “abstract syntax tree” or AST—which is then converted by the writer into the target format. The pandoc AST format is defined in the moduleText.Pandoc.Definition in thepandoc-types package.

A “filter” is a program that modifies the AST, between the reader and the writer.

INPUT --reader--> AST --filter--> AST --writer--> OUTPUT

Pandoc supports two kinds of filters:

• Lua filters use the Lua language to define transformations on the pandoc AST. They are described in aseparate document.

• JSON filters, described here, are pipes that read from standard input and write to standard output, consuming and producing a JSON representation of the pandoc AST:

                         source format                            ↓                         (pandoc)                            ↓                    JSON-formatted AST                            ↓                      (JSON filter)                            ↓                    JSON-formatted AST                            ↓                         (pandoc)                            ↓                      target format

Lua filters have a couple of advantages. They use a Lua interpreter that is embedded in pandoc, so you don’t need to have any external software installed. And they are usually faster than JSON filters. But if you wish to write your filter in a language other than Lua, you may prefer to use a JSON filter. JSON filters may be written in any programming language.

You can use a JSON filter directly in a pipeline:

pandoc -s input.txt -t json | \ pandoc-citeproc | \ pandoc -s -f json -o output.html

But it is more convenient to use the--filter option, which handles the plumbing automatically:

pandoc -s input.txt --filter pandoc-citeproc -o output.html

For a gentle introduction into writing your own filters, continue this guide. There’s also alist of third party filters on the wiki.

A simple example

Suppose you wanted to replace all level 2+ headings in a markdown document with regular paragraphs, with text in italics. How would you go about doing this?

A first thought would be to use regular expressions. Something like this:

perl -pe 's/^##+ (.*)$/\*\1\*/' source.txt

This should work most of the time. But don’t forget that ATX style headings can end with a sequence of#s that is not part of the heading text:

## My heading ##

And what if your document contains a line starting with## in an HTML comment or delimited code block?

<!--## This is just a comment-->~~~~### A third level heading in standard markdown~~~~

We don’t want to touchthese lines. Moreover, what about Setext style second-level heading?

A heading---------

We need to handle those too. Finally, can we be sure that adding asterisks to each side of our string will put it in italics? What if the string already contains asterisks around it? Then we’ll end up with bold text, which is not what we want. And what if it contains a regular unescaped asterisk?

How would you modify your regular expression to handle these cases? It would be hairy, to say the least.

A better approach is to let pandoc handle the parsing, and then modify the AST before the document is written. For this, we can use a filter.

To see what sort of AST is produced when pandoc parses our text, we can use pandoc’snative output format:

% cat test.txt## my headingtext with *italics*% pandoc -s -t native test.txtPandoc (Meta {unMeta = fromList []})[Header 2 ("my-heading",[],[]) [Str "My",Space,Str "heading"], Para [Str "text",Space,Str "with",Space,Emph [Str "italics"]] ]

APandoc document consists of aMeta block (containing metadata like title, authors, and date) and a list ofBlock elements. In this case, we have twoBlocks, aHeader and aPara. Each has as its content a list ofInline elements. For more details on the pandoc AST, see thehaddock documentation forText.Pandoc.Definition.

We can use Haskell to create a JSON filter that transforms this AST, replacing eachHeader block with level >= 2 with aPara with its contents wrapped inside anEmph inline:

#!/usr/bin/env runhaskell-- behead.hsimportText.Pandoc.JSONmain ::IO ()main= toJSONFilter beheadbehead ::Block->Blockbehead (Header n _ xs)| n>=2=Para [Emph xs]behead x= x

ThetoJSONFilter function does two things. First, it lifts thebehead function (which mapsBlock -> Block) onto a transformation of the entirePandoc AST, walking the AST and transforming each block. Second, it wraps thisPandoc -> Pandoc transformation with the necessary JSON serialization and deserialization, producing an executable that consumes JSON from stdin and produces JSON to stdout.

To use the filter, make it executable:

chmod +x behead.hs

and then

pandoc -f SOURCEFORMAT -t TARGETFORMAT --filter ./behead.hs

(It is also necessary thatpandoc-types be installed in the local package repository. To do this using cabal-install,cabal v2-update && cabal v2-install --lib pandoc-types --package-env ..)

Alternatively, we could compile the filter:

ghc -package-env=default --make behead.hspandoc -f SOURCEFORMAT -t TARGETFORMAT --filter ./behead

Note that if the filter is placed in the system PATH, then the initial./ is not needed. Note also that the command line can include multiple instances of--filter: the filters will be applied in sequence.

LaTeX for WordPress

Another easy example. WordPress blogs require a special format for LaTeX math. Instead of$e=mc^2$, you need:$LaTeX e=mc^2$. How can we convert a markdown document accordingly?

Again, it’s difficult to do the job reliably with regexes. A$ might be a regular currency indicator, or it might occur in a comment or code block or inline code span. We just want to find the$s that begin LaTeX math. If only we had a parser…

We do. Pandoc already extracts LaTeX math, so:

#!/usr/bin/env runhaskell-- wordpressify.hsimportText.Pandoc.JSONmain= toJSONFilter wordpressifywhere wordpressify (Math x y)=Math x ("LaTeX "++ y)        wordpressify x= x

Mission accomplished. (I’ve omitted type signatures here, just to show it can be done.)

But I don’t want to learn Haskell!

While it’s easiest to write pandoc filters in Haskell, it is fairly easy to write them in python using thepandocfilters package. The package is in PyPI and can be installed usingpip install pandocfilters oreasy_install pandocfilters.

Here’s our “beheading” filter in python:

#!/usr/bin/env python"""Pandoc filter to convert all level 2+ headings to paragraphs withemphasized text."""from pandocfiltersimport toJSONFilter, Emph, Paradef behead(key, value,format, meta):if key=='Header'and value[0]>=2:return Para([Emph(value[2])])if__name__=="__main__":  toJSONFilter(behead)

toJSONFilter(behead) walks the AST and applies thebehead action to each element. Ifbehead returns nothing, the node is unchanged; if it returns an object, the node is replaced; if it returns a list, the new list is spliced in.

Note that, although these parameters are not used in this example,format provides access to the target format, andmeta provides access to the document’s metadata.

There are many examples of python filters inthe pandocfilters repository.

For a more Pythonic alternative to pandocfilters, see thepanflute library. Don’t like Python? There are also ports of pandocfilters in

• PHP,
• perl,
• TypeScript/JavaScript via Node.js
  • pandoc-filter,
  • node-pandoc-filter,
• Groovy, and
• Ruby.

Starting with pandoc 2.0, pandoc includes built-in support for writing filters in lua. The lua interpreter is built in to pandoc, so a lua filter does not require any additional software to run. See thedocumentation on lua filters.

Include files

So none of our transforms have involved IO. How about a script that reads a markdown document, finds all the inline code blocks with attributeinclude, and replaces their contents with the contents of the file given?

#!/usr/bin/env runhaskell-- includes.hsimportText.Pandoc.JSONimportqualifiedData.Text.IOasTIOimportqualifiedData.TextasTdoInclude ::Block->IOBlockdoInclude cb@(CodeBlock (id, classes, namevals) contents)=caselookup (T.pack"include") namevalsofJust f->CodeBlock (id, classes, namevals)<$>                      TIO.readFile (T.unpack f)Nothing->return cbdoInclude x=return xmain ::IO ()main= toJSONFilter doInclude

Try this on the following:

Here's the pandoc README:~~~~ {include="README"}this will be replaced by contents of README~~~~

Removing links

What if we want to remove every link from a document, retaining the link’s text?

#!/usr/bin/env runhaskell-- delink.hsimportText.Pandoc.JSONmain= toJSONFilter delinkdelink ::Inline-> [Inline]delink (Link _ txt _)= txtdelink x= [x]

Note thatdelink can’t be a function of typeInline -> Inline, because the thing we want to replace the link with is not a singleInline element, but a list of them. So we makedelink a function from anInline element to a list ofInline elements.toJSONFilter can still lift this function to a transformation of typePandoc -> Pandoc.

A filter for ruby text

Finally, here’s a nice real-world example, developed on thepandoc-discuss list. Qubyte wrote:

I’m interested in using pandoc to turn my markdown notes on Japanese into nicely set HTML and (Xe)LaTeX. With HTML5, ruby (typically used to phonetically read chinese characters by placing text above or to the side) is standard, and support from browsers is emerging (Webkit based browsers appear to fully support it). For those browsers that don’t support it yet (notably Firefox) the feature falls back in a nice way by placing the phonetic reading inside brackets to the side of each Chinese character, which is suitable for other output formats too. As for (Xe)LaTeX, ruby is not an issue.

At the moment, I use inline HTML to achieve the result when the conversion is to HTML, but it’s ugly and uses a lot of keystrokes, for example

<ruby>ご<rt></rt>飯<rp>（</rp><rt>はん</rt><rp>）</rp></ruby>

sets ご飯 “gohan” with “han” spelt phonetically above the second character, or to the right of it in brackets if the browser does not support ruby. I’d like to have something more like

r[はん](飯)

or any keystroke saving convention would be welcome.

We came up with the following script, which uses the convention that a markdown link with a URL beginning with a hyphen is interpreted as ruby:

[はん](-飯)

{-# LANGUAGE OverloadedStrings #-}-- handleruby.hsimportText.Pandoc.JSONimportSystem.Environment (getArgs)importqualifiedData.TextasThandleRuby ::MaybeFormat->Inline->InlinehandleRuby (Just format) x@(Link attr [Str ruby] (src,_))=case T.uncons srcofJust ('-',kanji)| format==Format"html"->RawInline format$"<ruby>"<> kanji<>"<rp>(</rp><rt>"<> ruby<>"</rt><rp>)</rp></ruby>"| format==Format"latex"->RawInline format$"\\ruby{"<> kanji<>"}{"<> ruby<>"}"|otherwise->Str ruby    _-> xhandleRuby _ x= xmain ::IO ()main= toJSONFilter handleRuby

Note that, when a script is called using--filter, pandoc passes it the target format as the first argument. When a function’s first argument is of typeMaybe Format,toJSONFilter will automatically assign itJust the target format orNothing.

We compile our script:

# first, make sure pandoc-types is installed:cabal install --lib pandoc-types --package-env .ghc --make handleRuby

Then run it:

% pandoc -F ./handleRuby -t html[はん](-飯)^D<p><ruby>飯<rp>(</rp><rt>はん</rt><rp>)</rp></ruby></p>% pandoc -F ./handleRuby -t latex[はん](-飯)^D\ruby{飯}{はん}

Note: to use this to generate PDFs via LaTeX, you’ll need to use--pdf-engine=xelatex, specify amainfont that has the Japanese characters (e.g. “Noto Sans CJK JP”), and add\usepackage{ruby} to your template or header-includes.

Exercises

1. Put all the regular text in a markdown document in ALL CAPS (without touching text in URLs or link titles).

2. Remove all horizontal rules from a document.

3. Renumber all enumerated lists with roman numerals.

4. Replace each delimited code block with classdot with an image generated by runningdot -Tpng (from graphviz) on the contents of the code block.

5. Find all code blocks with classpython and run them using the python interpreter, printing the results to the console.

Technical details of JSON filters

A JSON filter is any program which can consume and produce a valid pandoc JSON document representation. This section describes the technical details surrounding the invocation of filters.

pandoc --filter demo --to=html