Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Table of Contents


CaveMark is a typesetting markdown parser. It is also the fastest pure-Python markdown parser!

If you don't know what's a typesetting parser read this . I don't know any typesetting markdown parser for Python. So maybe CaveMark is the only one. I know about Madoko, but its syntax sucks and I guess it's not a Python module (JavaScript). I never used Madoko.

This typesetting aspect makes CaveMark the easiest way to allow you write pretty documents for your blog/website/whatever, than, say, mistune or any other conventional markdown parser around.

E.g. by simply typing [myimage] shows a happy cat[mynote]. (assuming resources myimage and mynote are defined) you will conveniently get this beautifully formatted page, where figures, captions, footnotes, links, etc are all done automatically:

rendered output

Every single HTML code that CaveMark generates is configurable via self.frmt_* variables. There is no hard-coded HTML strings. E.g. while by default # This is a heading becomes <h1>This is a heading</h1>, you are free to change it to be, say, lol This is a heading rofl, or whatever.

For a production demonstration of what CaveMark can do, checkout this website; all its articles are written in CaveMark!


  • Headings with automatic indices.
  • Paragraphs.
  • Emphasized texts.
  • Strikethrough texts.
  • Lists.
  • Shortcuts.
  • Inline/block codes.
  • Inline/box resources (e.g. books, links, figures, quotes, theorems, definitions, etc).
  • Footnotes.
  • Bibliographies.
  • Simple, user-friendly, syntax (I think simplest among all typesetting parsers).
  • Fastest in class.
  • Supports CPython2, Cpython3, PyPy2, PyPy3.
  • Made with beard.


Speed is not the main selling point, but I think it shows that I didn't take too sloppy shortcuts while coding it. I am benchmarking against mistune coz it's the fastest in town. Others were too slow to even bother.

Since mistune and CaveMark have different syntax, this benchmark only tests things where they agree on their syntax. Even though, I think, CaveMark still does more. E.g. headings in CaveMark automatically also produce section identifiers. E.g:

# Heading

## Subheading

# Another heading

gives in mistune this:



<h1>Another heading<h1>

But gives in CaveMark:

<h1 id="sec_1"><a href="#sec_1">1.  Heading</a></h1>

<h2 id="sec_1.1"><a href="#sec_1.1">1.1.  Subheading</a></h2>

<h1 id="sec_2"><a href="#sec_2">2.  Another headinng</a></h1>

Results with CPython3 (repeated 3 times):

mistune : 18.177862088 seconds
cavemark: 9.386662586 seconds (1.9 times faster!)

mistune : 18.733004819999998 seconds
cavemark: 9.601156947 seconds (2.0 times faster!)

mistune : 18.762310276999997 seconds
cavemark: 9.647228555 seconds (1.9 times faster!)

Results with CPython2 (repeated 3 times):

mistune : 19.7261228561 seconds
cavemark: 9.48487520218 seconds (2.1 times faster!)

mistune : 19.4376080036 seconds
cavemark: 9.53958702087 seconds (2.0 times faster!)

mistune : 20.3399410248 seconds
cavemark: 9.61795091629 seconds (2.1 times faster!)

Results with PyPy3 (repeated 3 times):

mistune : 12.311563448000001 seconds
cavemark: 3.1916073270000003 seconds (3.9 times faster!)

mistune : 12.470256598999999 seconds
cavemark: 3.376128782 seconds (3.7 times faster!)

mistune : 13.637612451999999 seconds
cavemark: 3.126377384 seconds (4.4 times faster!)

Results with PyPy2 (repeated 3 times):

mistune : 8.99331712723 seconds
cavemark: 1.47040700912 seconds (6.1 times faster!)

mistune : 9.01434993744 seconds
cavemark: 1.45837283134 seconds (6.2 times faster!)

mistune : 9.00144791603 seconds
cavemark: 1.47606897354 seconds (6.1 times faster!)

You can find the full benchmark code in the benchmark directory.

In the future, I may do another benchmark where I test all the overlapping features, even when the syntax is not identical accross CaveMark and mistune. I plan to create two separate texts, both rendering exactly the same desired HTML document, but one in CaveMark's superior-yet-simpler typesetting syntax, and another in mistune's.


#[BOOKMARK] HEADING defines section headings, alongside its optional BOOKMARK. E.g.:

  • # heading becomes <h1 id='sec_1'><a href='#sec_1'>1.</a> heading</h1>.
  • #some_bookmark heading becomes <h1 id='some_bookmark'><a href='#some_bookmark'>1.</a>heading</h1>.

_ emphasizes texts. E.g. _this_ = this. ~~ to strike through text. E.g. ~~this~~ = this.

Ordered lists are defined by +, and the unordered by *. E.g.:

* this item is not orderd.
* another unordered item.
    + this item is nested and is ordered.
    + another ordered item.

The most interesting part of CaveMark is how it defines resources, and uses them to achieve typesetting. Below is an example where a figure is defined. This figure can then be cited later in the document (as shown later in this section).

image  : myimage
url    :
caption: caveman's cat looks happy
alt    : caveman's cat

Aside from image, many other resource types are defined, e.g. book, footnote, quote, theorem, etc. You can also easily add your own custom ones.

Citing a resource is by using []. E.g. [myimage] will cite the myimage resource we defined earlier. When a resource identifier, e.g. myimage is prefixed by !, in a citation box, e.g. [!myimage], then the resource myimage will not expand into its inline citation format. Using [!myimage] is useful when you want to only place the box/bibliography/footnote expansion of the resource, without placing the inline expansion.

You can also pass an optional data string to the citation procedure, when citing something. E.g. Recently, it was shown that [mylink:chickens can fly!], where mylink is a link resource that was defined earlier. This will then work similar to [chickens can fly!](https://...), except better and prettier.

Finally, a default resource type is implied if no specific resource identifier is specified in a citation. E.g. by default, [:some text] will be interpreted as a citation of the footnote with the automatically-generated identifier __1 (this ID is also configurable, where 1 keeps incrementing to keep each default identifiers unique):

footnote: __1
text    : some text

CaveMark also supports shortcuts. By default:

  • (c) = &copy;.
  • (tm) = &trade;.
  • (R) = &reg;.
  • " = &ldquo; .
  • '' = &rdquo; .
  • -- = &mdash; .
  • ... = &hellip;.

Basic Example

import cavemark

text_input = '''
image  : myimage
url    :
caption: Caveman's cat.
alt    : Caveman's cat.

book     : mybook
title    : The Book
authors  : Dude McDudeface
publisher: CavePress
year     : 2018

footnote: mynote
text    : She is 1 years old.

quotation: myquote
text     : Beardless men are entry-level trannis
author   : Dude McDudeface, 2018

# This is a _heading_

[myimage] shows a happy cat[mynote].

* A nice list.
* With items in it.
    + Can be nested.

    With, possibly, multiple paragraphs in it!

    Yep, another paragraph, with citations [mybook].

    + Another ordered item, but this one with ~~some struck-through text~~.

A paragraph could have `inline code like this`, or:

```#include <stdio.h>
int main(){
    printf("a code block like this!\\n");
    return 0;

How about citing [myquote]?

parser = cavemark.CaveMark()
html_output = parser.get_html()

with open('test001.html', 'w') as f:

If you open test.html using your browser, you will see something neat:

rendered output

CaveMark's Syntax Philosophy

I tried to avoid needless formatting features in CaveMark. E.g. most markdown parsers have several ways of doing the same thing. For example a heading could be defined by # This is a heading or:

This is a heading

The # method is superior, since it allows you to also specify heading's level by repeating #. E.g. ###### This is a heading defines a level 6 heading.

Therefore, CaveMark only supports the # method, and not the under-dashed one. I don't see a good reason to do it. Specially that text editors, such as vim, highlight headings texts accordingly either way.

Another example of avoiding needless features is that there is only one way to define emphasized texts. I don't see any reason to allow to offer several ways to emphasize texts. E.g. some markdown implementations offer *, **, _, __ to denote that a text is emphasized. I think this is needless. CaveMark only offers _ emphasize texts. E.g. _this_ becomes <em>this</em> in HTML by default. You can then style the em tags as you want, or completely replace the em tags by whatever you want.

I basically think that emphasization of texts is a semantic unit, and it is of one type. I see no case where it is good to emphasize texts by underlines in some paragraph, and then emphasize by bold in another paragraph. To me this makes no sense. IMO, if you wish to denote that a text is emphasized, then stick to the standard format for such a thing. In other words, varying the emphasization format across your paragraphs is like changing the font and color used in your section headings.

An example of where CaveMark reduces space of reserved keywords is that, in markdown, commonly, you cite a link by putting [some text](https://...) in your paragraphs. IMO this is inconvenient for two reasons:

  • Citing ugly URLs will clutter your paragraphs.
  • Too many keywords are reserved for the parser, namely [...] and (...). Do we really need to do this?

CaveMark solves those inconveniences by using the syntax [mylink:some text] instead, where mylink is a link resource defined elsewhere, e.g.:

url :

CaveMark's approach is superior because:

  • It reserves fewer keywords in practice. I.e. it only reserves [... : ...], so you can use (...) in your text without worrying about escaping them. The fact that : is reserved inside a [...] block is not worrying, because a resource ID is just an identifier (e.g. you don't need to use : in an identifier).
  • It keeps the ugly URLs away from your paragraphs. You can cite as many ugly URLs as you want, without being disturbed by damaging the readability of your markdown text.

If you can convince me that more features are needed, I will change my mind. But so far I have not seen any reason to justify having multiple ways of defining section headings or emphasized texts.


General Use

To use CaveMark:

  1. import cavemark.
  2. Initialize a CaveMark parsing object:
parser = cavemark.CaveMark()
  1. Pass CaveMark text text to the parser:
  1. Get HTML output
html = parser.get_html()

Flushing with parser.flush() is required to ensure you don't leave some leftover shit behind you.

Note: this will flush everything, including the body, footnotes and bibliographies. Sometimes, you wish to only flush one of these three. E.g. sin case you are parsing several articles, and you wish to only flush their HTML body and footnote, but not the bibliography, as you may want to gather the bibliography and flush it later after all the articles are parsed. To achieve this flexibility, you can pass the following arguments to parser.flush():

  • footnotes: Set it to True (default) to flush footnotes, or False to not flush footnotes yet (gather footnotes for a later flush).
  • bibliography: Set it to True (default) to flush the bibliography, or False to not flush the bibliography yet (gather bibliography items for a later flush).

E.g. if parser.flush(footnotes=False, bibliography=False), then only the HTML body is flushed.

  1. Use HTML html in your website however you please.

Making CaveMark Forget Stuff

In order to avoid needing to-reinitialize new CaveMark objects per document, as you may only wish to reset only simple aspect, such as resetting section heading indices, citation indices, etc. Why initialize a whole new CaveMark parser object, right?

Here is what you have:

  • parser.forget_resources()

    Forgets all resources that were found in text, unless resource=RESID is given then only that specific resource, RESID, will be forgotten.

  • parser.forget_citations()

    For any resource in text, forgets the that fact a resource was cited, unless resource=RESID is given, then only that specific resource, RESID, will have its citation forgotten.

    parser.forget_citations() differs against parser.forget_resources() in that the former forgets the fact that a resource was cited, while the latter forgets resource's existance/definition.

  • parser.forget_citation_counters()

    Reset counters of all cited resources, unless resource_type=RESTYPE is given then only resource indecies of that type RESTYPE will be reset.

    This is useful so when you wish to parse a new page, you don't need to kcontinue your citation indices from the previous page. E.g. if the last citation index was [100], and you start parsing a new page, you want to reset it so that your first cited resource is [1] instead of [101].

  • parser.forget_section_counters()

    Forgets the section heading ineces. E.g. if the last section was, say,, then calling this method will forget this fact, so that subsequently parsed section headings will start from the beginning, e.g. 1.. This is useful when you are about to parse a new unrelated page to what you were parsing previously.

Customizing CaveMark

Two ways:

  • During CaveMark parsing object initialization:

    parser = cavemark.CaveMark(option1=value1, option2=value2, ...)
  • After CaveMark parsing object initialization:

    # ... etc

    As you see, with this approach, calling parser.update() is required in order to recompile regular expressions, and other things, in order to ensure that all your changes are applied.

List of Options

  • resources

    Contains parsed resources in HTML form, as found in your passed text. This can be a convenient way for you to see all the resources that were parsed in the gtext that you have given to parser.parse(text).

    This can also be useful as a caching mechanism. E.g. for resources that you frequently use, you don't need to parse them each time into HTML. Instead, you can simply store parser.resources somewhere, and reload it parser.resources=old_parser_resources before parser.parse(text). This way, you won't need to include resources in raw format in text as the HTML-cached one is already supplied in parser.resources.

  • resource_keys_ignored

    A set containing resource keys that their corresponding value must not be parsed. E.g. url is one.

  • resource_counters

    A dictionary specifying counter names of various resources. If two entries have the same counter name, it means that they will share they same counter. E.g. by default link and book share the same counter:

    resource_counters = {
        'link'      :'counter_a', # sharing same counters
        'book'      :'counter_a', # sharing same counters
        'image'     :'counter_c',
        'note'      :'counter_d',
        'quotation' :'counter_e',
        'axiom'     :'counter_g',
        'theorem'   :'counter_h',
        'corollary' :'counter_i',
        'footnote'  :'counter_k',
  • escape

    Specifies the escaping sequence. By default, it's the backward slash.

  • code

    Defines opening/closing tags for code substrings, or any substring that you wish to ignore parsing on. By default:

    code = {
        '$$' : '$$',
        '$'  : '$',
        '`'  : '`',
        '```': '```',
  • code_inline

    Specifies whether a specific code substring (identified by its opening tag) should be always in a paragraph. By default:

    code_inline= {'$$', '$', '`'}
  • code_unescape

    Code substrings that unescapping should be enabled in them. By default:

    code_unescape = {'`', '```'}

    Note: the only unescapping that's enabled for them, which is the only one that makes sense, is unescapping strings that match their closing tag.

  • heading_offset

    Specifies an offset for all heading levels. By default:


    E.g. with the default offset of 1, if # heading is seen, then its level will be 2 instead of 1. E.g. if ## heading is seen, then its level will be 3 instead of 2.

  • shortcuts

    A dictionary for simple replacement of some characters by other charactters. By default:

    shortcuts = {
        '(c)'   : '&copy;', 
        '(tm)'  : '&trade;', 
        '(R)'   : '&reg;', 
        '"'     : '&ldquo;', 
        "''"    : '&rdquo;', 
        '---'   : '&mdash;', 
        '...'   : '&hellip;', 
  • cite_id_default

    If a resource citation lacks an identifier, then that cite_id_default will be used as its identifier prefix (the suffix is a counter to keep dthe identifiers unique). E.g. [:some text] lacks an identifier. By default:

    cite_id_default = '__'
  • cite_type_default

    The resource type for the cited resource that lacked a resource identifier. I.e. the resource type of resource ID __N. where N is some unique number, as set by cite_id_default. By default:

    cite_type_default = 'footnote'
  • cite_key_default

    The resource key for the cited resource that had optional data supplied. E.g. [RESID:some optional data]. This key depends on the resource type that the RESID belongs to. So, this is a dictionary that maps resource type against its default key that the optional data will be assigned to, By default:

    cite_key_default = {
        'link'      : 'text',
        'footnote'  : 'text',
  • frmt_cite_inline, frmt_cite_box, frmt_cite_data_default, frmt_cite_error_inline, frmt_cite_error_box, frmt_bibliography_prefix, frmt_bibliography_suffix, frmt_bibliography_item, frmt_bibliography_error_item, frmt_footnote_prefix, frmt_footnote_suffix, frmt_footnote_item, frmt_footnote_error_item, frmt_paragraph_prefix, frmt_paragraph_suffix, frmt_emph_prefix, frmt_emph_suffix, frmt_strike_prefix, frmt_strike_suffix, frmt_code_prefix, frmt_code_suffix, frmt_olist_prefix, frmt_olist_suffix, frmt_ulist_prefix, frmt_ulist_suffix, frmt_list_item_prefix, frmt_list_item_suffix, frmt_heading_prefix, frmt_heading_suffix

    All these are HTML formatting variables. They allow you to change every HTML code is generated by CaveMark.

    They are simple stuff. To see how they look like, just pring them. E.g. print(parser.frmt_cite_inline) to see the default value. I am sure anyone could figure out how to use them.


  • Tables.

Note 1: I currently don't need these features. My plan is to wait, until I happen to need them, then implement them. You are more than welcome to submit patches/pull requests. Alternatively, you can try to motivate me enough so that I implement them for you.


Typesetting markdown parser, that's also the fastest among all pure-Python markdown parsers!



No packages published