blob: 5e46f1dc38846ee9385399087ed557f8475ef587 [file] [log] [blame]
'''
Abbreviation Extension for Python-Markdown
==========================================
This extension adds abbreviation handling to Python-Markdown.
Simple Usage:
>>> import markdown
>>> text = """
... Some text with an ABBR and a REF. Ignore REFERENCE and ref.
...
... *[ABBR]: Abbreviation
... *[REF]: Abbreviation Reference
... """
>>> print markdown.markdown(text, ['abbr'])
<p>Some text with an <abbr title="Abbreviation">ABBR</abbr> and a <abbr title="Abbreviation Reference">REF</abbr>. Ignore REFERENCE and ref.</p>
Copyright 2007-2008
* [Waylan Limberg](http://achinghead.com/)
* [Seemant Kulleen](http://www.kulleen.org/)
'''
from __future__ import absolute_import
from __future__ import unicode_literals
from . import Extension
from ..preprocessors import Preprocessor
from ..inlinepatterns import Pattern
from ..util import etree
import re
# Global Vars
ABBR_REF_RE = re.compile(r'[*]\[(?P<abbr>[^\]]*)\][ ]?:\s*(?P<title>.*)')
class AbbrExtension(Extension):
""" Abbreviation Extension for Python-Markdown. """
def extendMarkdown(self, md, md_globals):
""" Insert AbbrPreprocessor before ReferencePreprocessor. """
md.preprocessors.add('abbr', AbbrPreprocessor(md), '<reference')
class AbbrPreprocessor(Preprocessor):
""" Abbreviation Preprocessor - parse text for abbr references. """
def run(self, lines):
'''
Find and remove all Abbreviation references from the text.
Each reference is set as a new AbbrPattern in the markdown instance.
'''
new_text = []
for line in lines:
m = ABBR_REF_RE.match(line)
if m:
abbr = m.group('abbr').strip()
title = m.group('title').strip()
self.markdown.inlinePatterns['abbr-%s'%abbr] = \
AbbrPattern(self._generate_pattern(abbr), title)
else:
new_text.append(line)
return new_text
def _generate_pattern(self, text):
'''
Given a string, returns an regex pattern to match that string.
'HTML' -> r'(?P<abbr>[H][T][M][L])'
Note: we force each char as a literal match (in brackets) as we don't
know what they will be beforehand.
'''
chars = list(text)
for i in range(len(chars)):
chars[i] = r'[%s]' % chars[i]
return r'(?P<abbr>\b%s\b)' % (r''.join(chars))
class AbbrPattern(Pattern):
""" Abbreviation inline pattern. """
def __init__(self, pattern, title):
super(AbbrPattern, self).__init__(pattern)
self.title = title
def handleMatch(self, m):
abbr = etree.Element('abbr')
abbr.text = m.group('abbr')
abbr.set('title', self.title)
return abbr
def makeExtension(configs=None):
return AbbrExtension(configs=configs)