Module Markup

Error-recovering streaming HTML and XML parsers and writers. is an HTML and XML parsing and serialization library. It:

The usage is straightforward. For example:

open Markup

(* Correct and pretty-print HTML. *)
channel stdin
|> parse_html |> signals |> pretty_print
|> write_html |> to_channel stdout

(* Show up to 10 XML well-formedness errors to the user. Stop after
   the 10th, without reading more input. *)
let report =
  let count = ref 0 in
  fun location error ->
    error |> Error.to_string ~location |> prerr_endline;
    count := !count + 1;
    if !count >= 10 then raise_notrace Exit

string "some xml" |> parse_xml ~report |> signals |> drain

(* Load HTML into a custom document tree data type. *)
type html = Text of string | Element of string * html list

file "some_file"
|> fst
|> parse_html
|> signals
|> tree
  ~text:(fun ss -> Text (String.concat "" ss))
  ~element:(fun (_, name) _ children -> Element (name, children))

The interface is centered around four functions. In pseudocode:

val parse_html : char stream   -> signal stream
val write_html : signal stream -> char stream
val parse_xml  : char stream   -> signal stream
val write_xml  : signal stream -> char stream

Most of the remaining functions create streams from, or write streams to, strings, files, and channels, or manipulate streams, such as next and the combinators map and fold.

Apart from this module, provides two other top-level modules:

Most of the interface of Markup_lwt is specified in signature ASYNCHRONOUS, which will be shared with a Markup_async module, should it be implemented. is developed on GitHub and distributed under the BSD license. This documentation is for version 0.8.1 of the library. Documentation for older versions can be found on the releases page.


type async
type sync

Phantom types for use with ('a, 's) stream in place of 's. See explanation below.

type ('a, 's) stream

Streams of elements of type 'a.

In simple usage, when using only this module Markup, the additional type parameter 's is always sync, and there is no need to consider it further.

However, if you are using Markup_lwt, you may create some async streams. The difference between the two is that next on a sync stream retrieves an element before next "returns," while next on an async stream might not retrieve an element until later. As a result, it is not safe to pass an async stream where a sync stream is required. The phantom types are used to make the type checker catch such errors at compile time.


The parsers recover from errors automatically. If that is sufficient, you can ignore this section. However, if you want stricter behavior, or need to debug parser output, use optional argument ?report of the parsers, and look in module Error.

type location = int * int

Line and column for parsing errors. Both numbers are one-based.

module Error : sig ... end

Error type and to_string function.


The parsers detect encodings automatically. If you need to specify an encoding, use optional argument ?encoding of the parsers, and look in module Encoding.

module Encoding : sig ... end

Common Internet encodings such as UTF-8 and UTF-16; also includes some less popular encodings that are sometimes used for XML.


type name = string * string

Expanded name: a namespace URI followed by a local name.

type xml_declaration = {
version : string;
encoding : string option;
standalone : bool option;

Representation of an XML declaration, i.e. <?xml version="1.0" encoding="utf-8"?>.

type doctype = {
doctype_name : string option;
public_identifier : string option;
system_identifier : string option;
raw_text : string option;
force_quirks : bool;

Representation of a document type declaration. The HTML parser fills in all fields besides raw_text. The XML parser reads declarations roughly, and fills only the raw_text field with the text found in the declaration.

type signal = [
| `Start_element of name * (name * string) list
| `End_element
| `Text of string list
| `Doctype of doctype
| `Xml of xml_declaration
| `PI of string * string
| `Comment of string

Parsing signals. The parsers emit them according to the following grammar:

doc     ::= `Xml? misc* `Doctype? misc* element misc*
misc    ::= `PI | `Comment
element ::= `Start_element content* `End_element
content ::= `Text | element | `PI | `Comment

As a result, emitted `Start_element and `End_element signals are always balanced, and, if there is an XML declaration, it is the first signal.

If parsing with ~context:`Document, the signal sequence will match the doc production until the first error. If parsing with ~context:`Fragment, it will match content*. If ~context is not specified, the parser will pick one of the two by examining the input.

As an example, if the XML parser is parsing

<?xml version="1.0"?><root>text<nested>more text</nested></root>

it will emit the signal sequence

`Xml {version = "1.0"; encoding = None; standalone = None}
`Start_element (("", "root"), [])
`Text ["text"]
`Start_element (("", "nested"), [])
`Text ["more text"]

The `Text signal carries a string list instead of a single string because on 32-bit platforms, OCaml strings cannot be larger than 16MB. In case the parsers encounter a very long sequence of text, one whose length exceeds about Sys.max_string_length / 2, they will emit a `Text signal with several strings.

type content_signal = [
| `Start_element of name * (name * string) list
| `End_element
| `Text of string list

A restriction of type signal to only elements and text, i.e. no comments, processing instructions, or declarations. This can be useful for pattern matching in applications that only care about the content and element structure of a document. See the helper content.

val signal_to_string : [< signal ] -> string

Provides a human-readable representation of signals for debugging.


type 's parser

An 's parser is a thin wrapper around a (signal, 's) stream that supports access to additional information that is not carried directly in the stream, such as source locations.

val signals : 's parser -> (signal's) stream

Converts a parser to its underlying signal stream.

val location : _ parser -> location

Evaluates to the location of the last signal emitted on the parser's signal stream. If no signals have yet been emitted, evaluates to (1, 1).


val parse_xml : ?⁠report:(location -> Error.t -> unit) -> ?⁠encoding:Encoding.t -> ?⁠namespace:(string -> string option) -> ?⁠entity:(string -> string option) -> ?⁠context:[< `Document | `Fragment ] -> (char, 's) stream -> 's parser

Creates a parser that converts an XML byte stream to a signal stream.

For simple usage, string "foo" |> parse_xml |> signals.

If ~report is provided, report is called for every error encountered. You may raise an exception in report, and it will propagate to the code reading the signal stream.

If ~encoding is not specified, the parser detects the input encoding automatically. Otherwise, the given encoding is used.

~namespace is called when the parser is unable to resolve a namespace prefix. If it evaluates to Some s, the parser maps the prefix to s. Otherwise, the parser reports `Bad_namespace.

~entity is called when the parser is unable to resolve an entity reference. If it evaluates to Some s, the parser inserts s into the text or attribute being parsed without any further parsing of s. s is assumed to be encoded in UTF-8. If entity evaluates to None instead, the parser reports `Bad_token. See xhtml_entity if you are parsing XHTML.

The meaning of ~context is described at signal, above.

val write_xml : ?⁠report:((signal * int) -> Error.t -> unit) -> ?⁠prefix:(string -> string option) -> ([< signal ]'s) stream -> (char, 's) stream

Converts an XML signal stream to a byte stream.

If ~report is provided, it is called for every error encountered. The first argument is a pair of the signal causing the error and its index in the signal stream. You may raise an exception in report, and it will propagate to the code reading the byte stream.

~prefix is called when the writer is unable to find a prefix in scope for a namespace URI. If it evaluates to Some s, the writer uses s for the URI. Otherwise, the writer reports `Bad_namespace.


val parse_html : ?⁠report:(location -> Error.t -> unit) -> ?⁠encoding:Encoding.t -> ?⁠context:[< `Document | `Fragment of string ] -> (char, 's) stream -> 's parser

Similar to parse_xml, but parses HTML with embedded SVG and MathML, never emits signals `Xml or `PI, and ~context has a different type on tag `Fragment.

For HTML fragments, you should specify the enclosing element, e.g. `Fragment "body". This is because, when parsing HTML, error recovery and the interpretation of text depend on the current element. For example, the text


parses differently in title elements than in p elements. In the former, it is parsed as foo</bar>, while in the latter, it is foo followed by a parse error due to unmatched tag </bar>. To get these behaviors, set ~context to `Fragment "title" and `Fragment "p", respectively.

If you use `Fragment "svg", the fragment is assumed to be SVG markup. Likewise, `Fragment "math" causes the parser to parse MathML markup.

If ~context is omitted, the parser guesses it from the input stream. For example, if the first signal would be `Doctype, the context is set to `Document, but if the first signal would be `Start_element "td", the context is set to `Fragment "tr". If the first signal would be `Start_element "g", the context is set to `Fragment "svg".

val write_html : ?⁠escape_attribute:(string -> string) -> ?⁠escape_text:(string -> string) -> ([< signal ]'s) stream -> (char, 's) stream

Similar to write_xml, but emits HTML5 instead of XML. If ~escape_attribute and/or ~escape_text are provided, they are used instead of default escaping functions.

Input sources

val string : string -> (char, sync) stream

Evaluates to a stream that retrieves successive bytes from the given string.

val buffer : Stdlib.Buffer.t -> (char, sync) stream

Evaluates to a stream that retrieves successive bytes from the given buffer. Be careful of changing the buffer while it is being iterated by the stream.

val channel : Stdlib.in_channel -> (char, sync) stream

Evaluates to a stream that retrieves bytes from the given channel. If the channel cannot be read, the next read of the stream results in raising Sys_error.

Note that this input source is synchronous because Pervasives.in_channel reads are blocking. For non-blocking channels, see Markup_lwt_unix.

val file : string -> (char, sync) stream * (unit -> unit)

file path opens the file at path, then evaluates to a pair s, close, where reading from stream s retrieves successive bytes from the file, and calling close () closes the file.

The file is closed automatically if s is read to completion, or if reading s raises an exception. It is not necessary to call close () in these cases.

If the file cannot be opened, raises Sys_error immediately. If the file cannot be read, reading the stream raises Sys_error.

val fn : (unit -> char option) -> (char, sync) stream

fn f is a stream that retrives bytes by calling f (). If the call results in Some c, the stream emits c. If the call results in None, the stream is considered to have ended.

This is actually an alias for stream, restricted to type char.

Output destinations

val to_string : (char, sync) stream -> string

Eagerly retrieves bytes from the given stream and assembles a string.

val to_buffer : (char, sync) stream -> Stdlib.Buffer.t

Eagerly retrieves bytes from the given stream and places them into a buffer.

val to_channel : Stdlib.out_channel -> (char, sync) stream -> unit

Eagerly retrieves bytes from the given stream and writes them to the given channel. If writing fails, raises Sys_error.

val to_file : string -> (char, sync) stream -> unit

Eagerly retrieves bytes from the given stream and writes them to the given file. If writing fails, or the file cannot be opened, raises Sys_error. Note that the file is truncated (cleared) before writing. If you wish to append to file, open it with the appropriate flags and use to_channel on the resulting channel.

Stream operations

val stream : (unit -> 'a option) -> ('async) stream

stream f creates a stream that repeatedly calls f (). Each time f () evaluates to Some v, the next item in the stream is v. The first time f () evaluates to None, the stream ends.

val next : ('async) stream -> 'a option

Retrieves the next item in the stream, if any, and removes it from the stream.

val peek : ('async) stream -> 'a option

Retrieves the next item in the stream, if any, but does not remove the item from the stream.

val transform : ('a -> 'b -> 'c list * 'a option) -> 'a -> ('b's) stream -> ('c's) stream

transform f init s lazily creates a stream by repeatedly applying f acc v, where acc is an accumulator whose initial value is init, and v is consecutive values of s. Each time, f acc v evaluates to a pair (vs, maybe_acc'). The values vs are added to the result stream. If maybe_acc' is Some acc', the accumulator is set to acc'. Otherwise, if maybe_acc' is None, the result stream ends.

val fold : ('a -> 'b -> 'a) -> 'a -> ('bsync) stream -> 'a

fold f init s eagerly folds over the items v, v', v'', ... of s, i.e. evaluates f (f (f init v) v') v''...

val map : ('a -> 'b) -> ('a's) stream -> ('b's) stream

map f s lazily applies f to each item of s, and produces the resulting stream.

val filter : ('a -> bool) -> ('a's) stream -> ('a's) stream

filter f s is s without the items for which f evaluates to false. filter is lazy.

val filter_map : ('a -> 'b option) -> ('a's) stream -> ('b's) stream

filter_map f s lazily applies f to each item v of s. If f v evaluates to Some v', the result stream has v'. If f v evaluates to None, no item corresponding to v appears in the result stream.

val iter : ('a -> unit) -> ('async) stream -> unit

iter f s eagerly applies f to each item of s, i.e. evaluates f v; f v'; f v''...

val drain : ('async) stream -> unit

drain s eagerly consumes s. This is useful for observing side effects, such as parsing errors, when you don't care about the parsing signals themselves. It is equivalent to iter ignore s.

val of_list : 'a list -> ('async) stream

Produces a (lazy) stream from the given list.

val to_list : ('async) stream -> 'a list

Eagerly converts the given stream to a list.


val content : ([< signal ]'s) stream -> (content_signal's) stream

Converts a signal stream into a content_signal stream by filtering out all signals besides `Start_element, `End_element, and `Text.

val tree : ?⁠text:(string list -> 'a) -> ?⁠element:(name -> (name * string) list -> 'a list -> 'a) -> ?⁠comment:(string -> 'a) -> ?⁠pi:(string -> string -> 'a) -> ?⁠xml:(xml_declaration -> 'a) -> ?⁠doctype:(doctype -> 'a) -> ([< signal ]sync) stream -> 'a option

This function's type signature may look intimidating, but it is actually easy to use. It is best introduced by example:

type my_dom = Text of string | Element of name * my_dom list

"<p>HTML5 is <em>easy</em> to parse"
|> string
|> parse_html
|> signals
|> tree
  ~text:(fun ss -> Text (String.concat "" ss))
  ~element:(fun (name, _) children -> Element (name, children))

results in the structure

Element ("p" [
  Text "HTML5 is ";
  Element ("em", [Text "easy"]);
  Text " to parse"])

Formally, tree assembles a tree data structure of type 'a from a signal stream. The stream is parsed according to the following grammar:

stream  ::= node*
node    ::= element | `Text | `Comment | `PI | `Xml | `Doctype
element ::= `Start_element node* `End_element

Each time trees matches a production of node, it calls the corresponding function to convert the node into your tree type 'a. For example, when trees matches `Text ss, it calls ~text ss, if ~text is supplied. Similarly, when trees matches element, it calls ~element name attributes children, if ~element is supplied.

See trees if the input stream might have multiple top-level trees. This function tree only retrieves the first one.

val trees : ?⁠text:(string list -> 'a) -> ?⁠element:(name -> (name * string) list -> 'a list -> 'a) -> ?⁠comment:(string -> 'a) -> ?⁠pi:(string -> string -> 'a) -> ?⁠xml:(xml_declaration -> 'a) -> ?⁠doctype:(doctype -> 'a) -> ([< signal ]'s) stream -> ('a's) stream

Like tree, but converts all top-level trees, not only the first one. The trees are emitted on the resulting stream, in the sequence that they appear in the input.

type 'a node = [
| `Element of name * (name * string) list * 'a list
| `Text of string
| `Doctype of doctype
| `Xml of xml_declaration
| `PI of string * string
| `Comment of string

See from_tree below.

val from_tree : ('a -> 'a node) -> 'a -> (signalsync) stream

Deconstructs tree data structures of type 'a into signal streams. The function argument is applied to each data structure node. For example,

type my_dom = Text of string | Element of string * my_dom list

let dom =
  Element ("p", [
    Text "HTML5 is ";
    Element ("em", [Text "easy"]);
    Text " to parse"])

dom |> from_tree (function
  | Text s -> `Text s
  | Element (name, children) -> `Element (("", name), [], children))

results in the signal stream

`Start_element (("", "p"), [])
`Text ["HTML5 is "]
`Start_element (("", "em"), [])
`Text ["easy"]
`Text " to parse"
val elements : (name -> (name * string) list -> bool) -> ([< signal ] as a, 's) stream -> (('a's) stream's) stream

elements f s scans the signal stream s for `Start_element (name, attributes) signals that satisfy f name attributes. Each such matching signal is the beginning of a substream that ends with the corresponding `End_element signal. The result of elements f s is the stream of these substreams.

Matches don't nest. If there is a matching element contained in another matching element, only the top one results in a substream.

Code using elements does not have to read each substream to completion, or at all. However, once the using code has tried to get the next substream, it should not try to read a previous one.

val text : ([< signal ]'s) stream -> (char, 's) stream

Extracts all the text in a signal stream by discarding all markup. For each `Text ss signal, the result stream has the bytes of the strings ss, and all other signals are ignored.

val trim : ([> content_signal ] as a, 's) stream -> ('a's) stream

Trims insignificant whitespace in an HTML signal stream. Whitespace around flow ("block") content does not matter, but whitespace in phrasing ("inline") content does. So, if the input stream is

  <em>foo</em> bar

passing it through Markup.trim will result in

<div><p><em>foo</em> bar</p></div>

Note that whitespace around the </em> tag was preserved.

val normalize_text : ([> `Text of string list ] as a, 's) stream -> ('a's) stream

Concatenates adjacent `Text signals, then eliminates all empty strings, then all `Text [] signals. Signals besides `Text are unaffected. Note that signal streams emitted by the parsers already have normalized text. This function is useful when you are inserting text into a signal stream after parsing, or generating streams from scratch, and would like to clean up the `Text signals.

val pretty_print : ([> content_signal ] as a, 's) stream -> ('a's) stream

Adjusts the whitespace in the `Text signals in the given stream so that the output appears nicely-indented when the stream is converted to bytes and written.

This function is aware of the significance of whitespace in HTML, so it avoids changing the whitespace in phrasing ("inline") content. For example, pretty printing


results in


Note that no whitespace was inserted around <em> and </em>, because doing so would create a word break that wasn't present in the original stream.

val html5 : ([< signal ]'s) stream -> (signal's) stream

Converts a signal stream into an HTML5 signal stream by stripping any document type declarations, XML declarations, and processing instructions, and prefixing the HTML5 doctype declaration. This is useful when converting between XHTML and HTML.

val xhtml : ?⁠dtd:[< `Strict_1_0 | `Transitional_1_0 | `Frameset_1_0 | `Strict_1_1 ] -> ([< signal ]'s) stream -> (signal's) stream

Similar to html5, but does not strip processing instructions, and prefixes an XHTML document type declaration and an XML declaration. The ~dtd argument specifies which DTD to refer to in the doctype declaration. The default is `Strict_1_1.

val xhtml_entity : string -> string option

Translates XHTML entities. This function is for use with the ~entity argument of parse_xml when parsing XHTML.

val strings_to_bytes : (string, 's) stream -> (char, 's) stream

strings_to_bytes s is the stream of all the bytes of all strings in s.

val compare_locations : location -> location -> int

Orders locations according to their appearance in an input stream, i.e. first by line, and then, for locations on the same line, by column.


module Ns : sig ... end

Common namespace URIs.

Asynchronous interface

module type ASYNCHRONOUS = sig ... end interface for monadic I/O libraries such as Lwt and Async.

Conformance status

The HTML parser seeks to implement section 8 of the HTML5 specification. That section describes a parser, part of a full-blown user agent, that is building up a DOM representation of an HTML document. is neither inherently part of a user agent, nor does it build up a DOM representation. With respect to section 8 of HTML5, is concerned with only the syntax. When that section requires that the user agent perform an action, emits enough information for a hypothetical user agent based on it to be able to decide to perform this action. Likewise, seeks to emit enough information for a hypothetical user agent to build up a conforming DOM.

The XML parser seeks to be a non-validating implementation of the XML and Namespaces in XML specifications.

This rest of this section lists known deviations from HTML5, XML, and Namespaces in XML. Some of these deviations are meant to be corrected in future versions of, while others will probably remain. The latter satisfy some or all of the following properties:

  • They require non-local adjustment, especially of past nodes. For example, adjusting the start signal of the root node mid-way through the signal stream is difficult for a one-pass parser.
  • They are minor. Users implementing less than a conforming browser typically don't care about them. They typically have to do with obscure error recovery. There are no deviations affecting the parsing of well-formed input.
  • They can easily be corrected by code written over that builds up a DOM or maintains other auxiliary data structures during parsing.

To be corrected:

  • XML: There is no attribute value normalization.
  • HTML: foster parenting is not implemented, because it requires non-local adjustments.
  • HTML: Quirks mode is not honored. This affects the interaction between automatic closing of p elements and opening of table elements.
  • HTML: The parser has non-standard recovery from unmatched closing form tags in some situations.
  • HTML: The parser ignores interactions between form and template.
  • HTML: The form translation for isindex is completely ignored. isindex is handled as an unknown element.

To remain:

  • HTML: Except when detecting encodings, the parser does not try to read <meta> tags for encoding declarations. The user of should read these, if necessary. They are part of the emitted signal stream.
  • HTML: noscript elements are always parsed, as are script elements. For conforming behavior, if the user of "supports scripts," the user should serialize the content of noscript to a `Text signal using write_html.
  • HTML: Elements such as title that belong in head, but are found between head and body, are not moved into head.
  • HTML: <html> tags found in the body do not have their attributes added to the `Start_element "html" signal emitted at the beginning of the document.