Pandoc   a universal document converter

Using the pandoc API

Pandoc can be used as a Haskell library, to write your own conversion tools or power a web application. This document offers an introduction to using the pandoc API.

Detailed API documentation at the level of individual functions and types is available at

Pandoc’s architecture

Pandoc is structured as a set of readers, which translate various input formats into an abstract syntax tree (the Pandoc AST) representing a structured document, and a set of writers, which render this AST into various input formats. Pictorially:

[input format] ==reader==> [Pandoc AST] ==writer==> [output format]

This architecture allows pandoc to perform M × N conversions with M readers and N writers.

The Pandoc AST is defined in the pandoc-types package. You should start by looking at the Haddock documentation for Text.Pandoc.Definition. As you’ll see, a Pandoc is composed of some metadata and a list of Blocks. There are various kinds of Block, including Para (paragraph), Header (section heading), and BlockQuote. Some of the Blocks (like BlockQuote) contain lists of Blocks, while others (like Para) contain lists of Inlines, and still others (like CodeBlock) contain plain text or nothing. Inlines are the basic elements of paragraphs. The distinction between Block and Inline in the type system makes it impossible to represent, for example, a link (Inline) whose link text is a block quote (Block). This expressive limitation is mostly a help rather than a hindrance, since many of the formats pandoc supports have similar limitations.

The best way to explore the pandoc AST is to use pandoc -t native, which will display the AST correspoding to some Markdown input:

% echo -e "1. *foo*\n2. bar" | pandoc -t native
[OrderedList (1,Decimal,Period)
 [[Plain [Emph [Str "foo"]]]
 ,[Plain [Str "bar"]]]]

A simple example

Here is a simple example of the use of a pandoc reader and writer to perform a conversion:

import Text.Pandoc
import qualified Data.Text as T
import qualified Data.Text.IO as TIO
main :: IO ()
main = do
result <- runIO $ do
doc <- readMarkdown def (T.pack "[testing](url)")
writeRST def doc
rst <- handleError result
TIO.putStrLn rst

Some notes:

  1. The first part constructs a conversion pipeline: the input string is passed to readMarkdown, and the resulting Pandoc AST (doc) is then rendered by writeRST. The conversion pipeline is “run” by runIO—more on that below.

  2. result has the type Either PandocError Text. We could pattern-match on this manually, but it’s simpler in this context to use the handleError function from Text.Pandoc.Error. This exits with an appropriate error code and message if the value is a Left, and returns the Text if the value is a Right.

The PandocMonad class

Let’s look at the types of readMarkdown and writeRST:

readMarkdown :: PandocMonad m => ReaderOptions -> Text -> m Pandoc
writeRST :: PandocMonad m => WriterOptions -> Pandoc -> m Text

The PandocMonad m => part is a typeclass constraint. It says that readMarkdown and writeRST define computations that can be used in any instance of the PandocMonad type class. PandocMonad is defined in the module Text.Pandoc.Class.

Two instances of PandocMonad are provided: PandocIO and PandocPure. The difference is that computations run in PandocIO are allowed to do IO (for example, read a file), while computations in PandocPure are free of any side effects. PandocPure is useful for sandboxed environments, when you want to prevent users from doing anything malicious. To run the conversion in PandocIO, use runIO (as above). To run it in PandocPure, use runPure.

As you can see from the Haddocks, Text.Pandoc.Class exports many auxiliary functions that can be used in any instance of PandocMonad. For example:

-- | Get the verbosity level.
getVerbosity :: PandocMonad m => m Verbosity
-- | Set the verbosity level.
setVerbosity :: PandocMonad m => Verbosity -> m ()
-- Get the accomulated log messages (in temporal order).
getLog :: PandocMonad m => m [LogMessage]
getLog = reverse <$> getsCommonState stLog
-- | Log a message using 'logOutput'. Note that 'logOutput' is
-- called only if the verbosity level exceeds the level of the
-- message, but the message is added to the list of log messages
-- that will be retrieved by 'getLog' regardless of its verbosity level.
report :: PandocMonad m => LogMessage -> m ()
-- | Fetch an image or other item from the local filesystem or the net.
-- Returns raw content and maybe mime type.
fetchItem :: PandocMonad m
=> String
-> m (B.ByteString, Maybe MimeType)
-- Set the resource path searched by 'fetchItem'.
setResourcePath :: PandocMonad m => [FilePath] -> m ()

If we wanted more verbose informational messages during the conversion we defined in the previous section, we could do this:

result <- runIO $ do
setVerbosity INFO
doc <- readMarkdown def (T.pack "[testing](url)")
writeRST def doc

Note that PandocIO is an instance of MonadIO, so you can use liftIO to perform arbitrary IO operations inside a pandoc conversion chain.


The first argument of each reader or writer is for options controlling the behavior of the reader or writer: ReaderOptions for readers and WriterOptions for writers. These are defined in Text.Pandoc.Options. It is a good idea to study these options to see what can be adjusted.

def (from Data.Default) denotes a default value for each kind of option. (You can also use defaultWriterOptions and defaultReaderOptions.) Generally you’ll want to use the defaults and modify them only when needed, for example:

writeRST def{ writerReferenceLinks = True }

Some particularly important options to know about:

  1. writerTemplate: By default, this is Nothing, which means that a document fragment will be produced. If you want a full document, you need to specify Just template, where template is a String containing the template’s contents (not the path).

  2. readerExtensions and writerExtensions: These specify the extensions to be used in parsing and rendering. Extensions are defined in Text.Pandoc.Extensions.


Sometimes it’s useful to construct a Pandoc document programatically. To make this easier we provide the module Text.Pandoc.Builder pandoc-types.

Because concatenating lists is slow, we use special types Inlines and Blocks that wrap a Sequence of Inline and Block elements. These are instances of the Monoid typeclass and can easily be concatenated:

import Text.Pandoc.Builder
mydoc :: Pandoc
mydoc = doc $ header 1 (text "Hello!")
<> para (emph (text "hello world") <> text ".")
main :: IO ()
main = print mydoc

If you use the OverloadedStrings pragma, you can simplify this further:

mydoc = doc $ header 1 "Hello!"
<> para (emph "hello world" <> ".")

Here’s a more realistic example. Suppose your boss says: write me a letter in Word listing all the filling stations in Chicago that take the Voyager card. You find some JSON data in this format (fuel.json):

[ {
"state" : "IL",
"city" : "Chicago",
"fuel_type_code" : "CNG",
"zip" : "60607",
"station_name" : "Clean Energy - Yellow Cab",
"cards_accepted" : "A D M V Voyager Wright_Exp CleanEnergy",
"street_address" : "540 W Grenshaw"
}, ...

And then use aeson and pandoc to parse the JSON and create the Word document:

{-# LANGUAGE OverloadedStrings #-}
import Text.Pandoc.Builder
import Text.Pandoc
import Data.Monoid ((<>), mempty, mconcat)
import Data.Aeson
import Control.Applicative
import Control.Monad (mzero)
import qualified Data.ByteString.Lazy as BL
import qualified Data.Text as T
import Data.List (intersperse)
data Station = Station{
address :: String
, name :: String
, cardsAccepted :: [String]
} deriving Show
instance FromJSON Station where
parseJSON (Object v) = Station <$>
v .: "street_address" <*>
v .: "station_name" <*>
(words <$> (v .:? "cards_accepted" .!= ""))
parseJSON _ = mzero
createLetter :: [Station] -> Pandoc
createLetter stations = doc $
para "Dear Boss:" <>
para "Here are the CNG stations that accept Voyager cards:" <>
simpleTable [plain "Station", plain "Address", plain "Cards accepted"]
(map stationToRow stations) <>
para "Your loyal servant," <>
plain (image "JohnHancock.png" "" mempty)
stationToRow station =
[ plain (text $ name station)
, plain (text $ address station)
, plain (mconcat $ intersperse linebreak
$ map text $ cardsAccepted station)
main :: IO ()
main = do
json <- BL.readFile "fuel.json"
let letter = case decode json of
Just stations -> createLetter [s | s <- stations,
"Voyager" `elem` cardsAccepted s]
Nothing -> error "Could not decode JSON"
docx <- runIO (writeDocx def letter) >>= handleError
BL.writeFile "letter.docx" docx
putStrLn "Created letter.docx"

Voila! You’ve written the letter without using Word and without looking at the data.

Data files

Pandoc has a number of data files, which can be found in the data/ subdirectory of the repository. These are installed with pandoc (or, if pandoc was compiled with the embed_data_files flag, they are embedded in the binary). You can retrieve data files using readDataFile from Text.Pandoc.Class. readDataFile will first look for the file in the “user data directory” (setUserDataDir, getUserDataDir), and if it is not found there, it will return the default installed with the system. To force the use of the default, setUserDataDir Nothing.


Pandoc has its own template system, described in the User’s Guide. To retrieve the default template for a system, use getDefaultTemplate from Text.Pandoc.Templates. Note that this looks first in the templates subdirectory of the user data directory, allowing users to override the system defaults. If you want to disable this behavior, use setUserDataDir Nothing.

To render a template, use renderTemplate', which takes two arguments, a template (String) and a context (any instance of ToJSON). If you want to create a context from the metadata part of a Pandoc document, use metaToJSON' from Text.Pandoc.Writers.Shared. If you also want to incorporate values from variables, use metaToJSON instead, and make sure writerVariables is set in WriterOptions.

Handling errors and warnings

runIO and runPure return an Either PandocError a. All errors raised in running a PandocMonad computation will be trapped and returned as a Left value, so they can be handled by the calling program. To see the constructors for PandocError, see the documentation for Text.Pandoc.Error.

To raise a PandocError from inside a PandocMonad computation, use throwError.

In addition to errors, which stop execution of the conversion pipeline, one can generate informational messages. Use report from Text.Pandoc.Class to issue a LogMessage. For a list of cosntructors for LogMessage, see Text.Pandoc.Logging. Note that each type of log message is associated with a verbosity level. The verbosity level (setVerbosity/getVerbosity) determines whether the report will be printed to stderr (when running in PandocIO), but regardless of verbosity level, all reported messages are stored internally and may be retrieved using getLog.

Walking the AST

It is often useful to walk the Pandoc AST either to extract information (e.g., what are all the URLs linked to in this document?, do all the code samples compile?) or to transform a document (e.g., increase the level of every section header, remove emphasis, or replace specially marked code blocks with images). To make this easier and more efficient, pandoc-types includes a module Text.Pandoc.Walk.

Here’s the essential documentation:

class Walkable a b where
-- | @walk f x@ walks the structure @x@ (bottom up) and replaces every
-- occurrence of an @a@ with the result of applying @f@ to it.
walk :: (a -> a) -> b -> b
walk f = runIdentity . walkM (return . f)
-- | A monadic version of 'walk'.
walkM :: (Monad m, Functor m) => (a -> m a) -> b -> m b
-- | @query f x@ walks the structure @x@ (bottom up) and applies @f@
-- to every @a@, appending the results.
query :: Monoid c => (a -> c) -> b -> c

Walkable instances are defined for most combinations of Pandoc types. For example, the Walkable Inline Block instance allows you to take a function Inline -> Inline and apply it over every inline in a Block. And Walkable [Inline] Pandoc allows you to take a function [Inline] -> [Inline] and apply it over every maximal list of Inlines in a Pandoc.

Here’s a simple example of a function that promotes the levels of headers:

promoteHeaderLevels :: Pandoc -> Pandoc
promoteHeaderLevels = walk promote
where promote :: Block -> Block
promote (Header lev attr ils) = Header (lev + 1) attr ils
promote x = x

walkM is a monadic version of walk; it can be used, for example, when you need your transformations to perform IO operations, use PandocMonad operations, or update internal state. Here’s an example using the State monad to add unique identifiers to each code block:

addCodeIdentifiers :: Pandoc -> Pandoc
addCodeIdentifiers doc = evalState (walkM addCodeId doc) 1
where addCodeId :: Block -> State Int Block
addCodeId (CodeBlock (_,classes,kvs) code) = do
curId <- get
put (curId + 1)
return $ CodeBlock (show curId,classes,kvs) code
addCodeId x = return x

query is used to collect information from the AST. Its argument is a query function that produces a result in some monoidal type (e.g. a list). The results are concatenated together. Here’s an example that returns a list of the URLs linked to in a document:

listURLs :: Pandoc -> [String]
listURLs = query urls
where urls (Link _ _ (src, _)) = [src]
urls _ = []

Creating a front-end

All of the functionality of the command-line program pandoc has been abstracted out in convertWithOpts in the module Text.Pandoc.App. Creating a GUI front-end for pandoc is thus just a matter of populating the Opts structure and calling this function.

Notes on using pandoc in web applications

  1. Pandoc’s parsers can exhibit pathological behavior on some inputs. So it is always a good idea to wrap uses of pandoc in a timeout function (e.g. System.Timeout.timeout from base) to prevent DOS attacks.

  2. If pandoc generates HTML from untrusted user input, it is always a good idea to filter the generated HTML through a sanitizer (such as xss-sanitize) to avoid security problems.

  3. Using runPure rather than runIO will ensure that pandoc’s functions perform no IO operations (e.g. writing files). If some resources need to be made available, a “fake environment” is provided inside the state available to runPure (see PureState and its associated functions in Text.Pandoc.Class). It is also possible to write a custom instance of PandocMonad that, for example, makes wiki resources available as files in the fake environment, while isolating pandoc from the rest of the system.