Abstract

When plaintext hasn't been expressive enough for inline documentation, Python programmers have sought out a format for docstrings. This PEP proposes that the reStructuredText markup [5] be adopted as a standard markup format for structured plaintext documentation in Python docstrings, and for PEPs and ancillary documents as well. reStructuredText is a rich and extensible yet easy-to-read, what-you-see-is-what-you-get plaintext markup syntax.

Only the low-level syntax of docstrings is addressed here. This PEP is not concerned with docstring semantics or processing at all (see PEP 256 for a "Road Map to the Docstring PEPs"). Nor is it an attempt to deprecate pure plaintext docstrings, which are always going to be legitimate. The reStructuredText markup is an alternative for those who want more expressive docstrings.

Benefits

Programmers are by nature a lazy breed. We reuse code with functions, classes, modules, and subsystems. Through its docstring syntax, Python allows us to document our code from within. The "holy grail" of the Python Documentation Special Interest Group (Doc-SIG [6]) has been a markup syntax and toolset to allow auto-documentation, where the docstrings of Python systems can be extracted in context and processed into useful, high-quality documentation for multiple purposes.

Document markup languages have three groups of customers: the authors who write the documents, the software systems that process the data, and the readers, who are the final consumers and the most important group. Most markups are designed for the authors and software systems; readers are only meant to see the processed form, either on paper or via browser software. ReStructuredText is different: it is intended to be easily readable in source form, without prior knowledge of the markup. ReStructuredText is entirely readable in plaintext format, and many of the markup forms match common usage (e.g., *emphasis*), so it reads quite naturally. Yet it is rich enough to produce complex documents, and extensible so that there are few limits. Of course, to write reStructuredText documents some prior knowledge is required.

The markup offers functionality and expressivity, while maintaining easy readability in the source text. The processed form (HTML etc.) makes it all accessible to readers: inline live hyperlinks; live links to and from footnotes; automatic tables of contents (with live links!); tables; images for diagrams etc.; pleasant, readable styled text.

The reStructuredText parser is available now, part of the Docutils [24] project. Standalone reStructuredText documents and PEPs can be converted to HTML; other output format writers are being worked on and will become available over time. Work is progressing on a Python source "Reader" which will implement auto-documentation from docstrings. Authors of existing auto-documentation tools are encouraged to integrate the reStructuredText parser into their projects, or better yet, to join forces to produce a world-class toolset for the Python standard library.

Tools will become available in the near future, which will allow programmers to generate HTML for online help, XML for multiple purposes, and eventually PDF, DocBook, and LaTeX for printed documentation, essentially "for free" from the existing docstrings. The adoption of a standard will, at the very least, benefit docstring processing tools by preventing further "reinventing the wheel".

Eventually PyDoc, the one existing standard auto-documentation tool, could have reStructuredText support added. In the interim it will have no problem with reStructuredText markup, since it treats all docstrings as preformatted plaintext.

Goals

These are the generally accepted goals for a docstring format, as discussed in the Doc-SIG:

1. It must be readable in source form by the casual observer.
2. It must be easy to type with any standard text editor.
3. It must not need to contain information which can be deduced from parsing the module.
4. It must contain sufficient information (structure) so it can be converted to any reasonable markup format.
5. It must be possible to write a module's entire documentation in docstrings, without feeling hampered by the markup language.

reStructuredText meets and exceeds all of these goals, and sets its own goals as well, even more stringent. See Docstring-Significant Features below.

The goals of this PEP are as follows:

1. To establish reStructuredText as a standard structured plaintext format for docstrings (inline documentation of Python modules and packages), PEPs, README-type files and other standalone documents. "Accepted" status will be sought through Python community consensus and eventual BDFL pronouncement.

   Please note that reStructuredText is being proposed as a standard, not the only standard. Its use will be entirely optional. Those who don't want to use it need not.

2. To solicit and address any related concerns raised by the Python community.

3. To encourage community support. As long as multiple competing markups are out there, the development community remains fractured. Once a standard exists, people will start to use it, and momentum will inevitably gather.

4. To consolidate efforts from related auto-documentation projects. It is hoped that interested developers will join forces and work on a joint/merged/common implementation.

Once reStructuredText is a Python standard, effort can be focused on tools instead of arguing for a standard. Python needs a standard set of documentation tools.

With regard to PEPs, one or both of the following strategies may be applied:

1. Keep the existing PEP section structure constructs (one-line section headers, indented body text). Subsections can either be forbidden, or supported with reStructuredText-style underlined headers in the indented body text.
2. Replace the PEP section structure constructs with the reStructuredText syntax. Section headers will require underlines, subsections will be supported out of the box, and body text need not be indented (except for block quotes).

Strategy (b) is recommended, and its implementation is complete.

Support for RFC 2822 headers has been added to the reStructuredText parser for PEPs (unambiguous given a specific context: the first contiguous block of the document). It may be desired to concretely specify what over/underline styles are allowed for PEP section headers, for uniformity.

Rationale

The lack of a standard syntax for docstrings has hampered the development of standard tools for extracting and converting docstrings into documentation in standard formats (e.g., HTML, DocBook, TeX). There have been a number of proposed markup formats and variations, and many tools tied to these proposals, but without a standard docstring format they have failed to gain a strong following and/or floundered half-finished.

Throughout the existence of the Doc-SIG, consensus on a single standard docstring format has never been reached. A lightweight, implicit markup has been sought, for the following reasons (among others):

1. Docstrings written within Python code are available from within the interactive interpreter, and can be "print"ed. Thus the use of plaintext for easy readability.
2. Programmers want to add structure to their docstrings, without sacrificing raw docstring readability. Unadorned plaintext cannot be transformed ("up-translated") into useful structured formats.
3. Explicit markup (like XML or TeX) is widely considered unreadable by the uninitiated.
4. Implicit markup is aesthetically compatible with the clean and minimalist Python syntax.

Many alternative markups for docstrings have been proposed on the Doc-SIG over the years; a representative sample is listed below. Each is briefly analyzed in terms of the goals stated above. Please note that this is not intended to be an exclusive list of all existing markup systems; there are many other markups (Texinfo, Doxygen, TIM, YODL, AFT, ...) which are not mentioned.

• XML [7], SGML [8], DocBook [9], HTML [10], XHTML [11]

  XML and SGML are explicit, well-formed meta-languages suitable for all kinds of documentation. XML is a variant of SGML. They are best used behind the scenes, because to untrained eyes they are verbose, difficult to type, and too cluttered to read comfortably as source. DocBook, HTML, and XHTML are all applications of SGML and/or XML, and all share the same basic syntax and the same shortcomings.

• TeX [12]

  TeX is similar to XML/SGML in that it's explicit, but not very easy to write, and not easy for the uninitiated to read.

• Perl POD [13]

  Most Perl modules are documented in a format called POD (Plain Old Documentation). This is an easy-to-type, very low level format with strong integration with the Perl parser. Many tools exist to turn POD documentation into other formats: info, HTML and man pages, among others. However, the POD syntax takes after Perl itself in terms of readability.

• JavaDoc [14]

  Special comments before Java classes and functions serve to document the code. A program to extract these, and turn them into HTML documentation is called javadoc, and is part of the standard Java distribution. However, JavaDoc has a very intimate relationship with HTML, using HTML tags for most markup. Thus it shares the readability problems of HTML.

• Setext [15], StructuredText [16]

  Early on, variants of Setext (Structure Enhanced Text), including Zope Corp's StructuredText, were proposed for Python docstring formatting. Hereafter these variants will collectively be called "STexts". STexts have the advantage of being easy to read without special knowledge, and relatively easy to write.

  Although used by some (including in most existing Python auto-documentation tools), until now STexts have failed to become standard because:

  • STexts have been incomplete. Lacking "essential" constructs that people want to use in their docstrings, STexts are rendered less than ideal. Note that these "essential" constructs are not universal; everyone has their own requirements.
  • STexts have been sometimes surprising. Bits of text are unexpectedly interpreted as being marked up, leading to user frustration.
  • SText implementations have been buggy.
  • Most STexts have have had no formal specification except for the implementation itself. A buggy implementation meant a buggy spec, and vice-versa.
  • There has been no mechanism to get around the SText markup rules when a markup character is used in a non-markup context. In other words, no way to escape markup.

Proponents of implicit STexts have vigorously opposed proposals for explicit markup (XML, HTML, TeX, POD, etc.), and the debates have continued off and on since 1996 or earlier.

reStructuredText is a complete revision and reinterpretation of the SText idea, addressing all of the problems listed above.

Specification

The specification and user documentaton for reStructuredText is quite extensive. Rather than repeating or summarizing it all here, links to the originals are provided.

Please first take a look at A ReStructuredText Primer [17], a short and gentle introduction. The Quick reStructuredText [18] user reference quickly summarizes all of the markup constructs. For complete and extensive details, please refer to the following documents:

• An Introduction to reStructuredText [19]
• reStructuredText Markup Specification [20]
• reStructuredText Directives [21]

In addition, Problems With StructuredText [22] explains many markup decisions made with regards to StructuredText, and A Record of reStructuredText Syntax Alternatives [23] records markup decisions made independently.

Docstring-Significant Features

• A markup escaping mechanism.

  Backslashes (\) are used to escape markup characters when needed for non-markup purposes. However, the inline markup recognition rules have been constructed in order to minimize the need for backslash-escapes. For example, although asterisks are used for emphasis, in non-markup contexts such as "*" or "(*)" or "x * y", the asterisks are not interpreted as markup and are left unchanged. For many non-markup uses of backslashes (e.g., describing regular expressions), inline literals or literal blocks are applicable; see the next item.

• Markup to include Python source code and Python interactive sessions: inline literals, literal blocks, and doctest blocks.

  Inline literals use double-backquotes to indicate program I/O or code snippets. No markup interpretation (including backslash-escape [\] interpretation) is done within inline literals.

  Literal blocks (block-level literal text, such as code excerpts or ASCII graphics) are indented, and indicated with a double-colon ("::") at the end of the preceding paragraph (right here -->):

  if literal_block:
      text = 'is left as-is'
      spaces_and_linebreaks = 'are preserved'
      markup_processing = None
  

  Doctest blocks begin with ">>> " and end with a blank line. Neither indentation nor literal block double-colons are required. For example:

  Here's a doctest block:
  
  >>> print 'Python-specific usage examples; begun with ">>>"'
  Python-specific usage examples; begun with ">>>"
  >>> print '(cut and pasted from interactive sessions)'
  (cut and pasted from interactive sessions)
  

• Markup that isolates a Python identifier: interpreted text.

  Text enclosed in single backquotes is recognized as "interpreted text", whose interpretation is application-dependent. In the context of a Python docstring, the default interpretation of interpreted text is as Python identifiers. The text will be marked up with a hyperlink connected to the documentation for the identifier given. Lookup rules are the same as in Python itself: LGB namespace lookups (local, global, builtin). The "role" of the interpreted text (identifying a class, module, function, etc.) is determined implicitly from the namespace lookup. For example:

  class Keeper(Storer):
  
      """
      Keep data fresher longer.
  
      Extend `Storer`.  Class attribute `instances` keeps track
      of the number of `Keeper` objects instantiated.
      """
  
      instances = 0
      """How many `Keeper` objects are there?"""
  
      def __init__(self):
          """
          Extend `Storer.__init__()` to keep track of
          instances.  Keep count in `self.instances` and data
          in `self.data`.
          """
          Storer.__init__(self)
          self.instances += 1
  
          self.data = []
          """Store data in a list, most recent last."""
  
      def storedata(self, data):
          """
          Extend `Storer.storedata()`; append new `data` to a
          list (in `self.data`).
          """
          self.data = data
  

  Each piece of interpreted text is looked up according to the local namespace of the block containing its docstring.

• Markup that isolates a Python identifier and specifies its type: interpreted text with roles.

  Although the Python source context reader is designed not to require explicit roles, they may be used. To classify identifiers explicitly, the role is given along with the identifier in either prefix or suffix form:

  Use :method:`Keeper.storedata` to store the object's data in
  `Keeper.data`:instance_attribute:.
  

  The syntax chosen for roles is verbose, but necessarily so (if anyone has a better alternative, please post it to the Doc-SIG [6]). The intention of the markup is that there should be little need to use explicit roles; their use is to be kept to an absolute minimum.

• Markup for "tagged lists" or "label lists": field lists.

  Field lists represent a mapping from field name to field body. These are mostly used for extension syntax, such as "bibliographic field lists" (representing document metadata such as author, date, and version) and extension attributes for directives (see below). They may be used to implement methodologies (docstring semantics), such as identifying parameters, exceptions raised, etc.; such usage is beyond the scope of this PEP.

  A modified RFC 2822 syntax is used, with a colon before as well as after the field name. Field bodies are more versatile as well; they may contain multiple field bodies (even nested field lists). For example:

  :Date: 2002-03-22
  :Version: 1
  :Authors:
      - Me
      - Myself
      - I
  

  Standard RFC 2822 header syntax cannot be used for this construct because it is ambiguous. A word followed by a colon at the beginning of a line is common in written text.

• Markup extensibility: directives and substitutions.

  Directives are used as an extension mechanism for reStructuredText, a way of adding support for new block-level constructs without adding new syntax. Directives for images, admonitions (note, caution, etc.), and tables of contents generation (among others) have been implemented. For example, here's how to place an image:

  .. image:: mylogo.png
  

  Substitution definitions allow the power and flexibility of block-level directives to be shared by inline text. For example:

  The |biohazard| symbol must be used on containers used to
  dispose of medical waste.
  
  .. |biohazard| image:: biohazard.png
  

• Section structure markup.

  Section headers in reStructuredText use adornment via underlines (and possibly overlines) rather than indentation. For example:

  This is a Section Title
  =======================
  
  This is a Subsection Title
  --------------------------
  
  This paragraph is in the subsection.
  
  This is Another Section Title
  =============================
  
  This paragraph is in the second section.
  

Questions & Answers

1. Is reStructuredText rich enough?

   Yes, it is for most people. If it lacks some construct that is required for a specific application, it can be added via the directive mechanism. If a useful and common construct has been overlooked and a suitably readable syntax can be found, it can be added to the specification and parser.

2. Is reStructuredText too rich?

   For specific applications or individuals, perhaps. In general, no.

   Since the very beginning, whenever a docstring markup syntax has been proposed on the Doc-SIG [6], someone has complained about the lack of support for some construct or other. The reply was often something like, "These are docstrings we're talking about, and docstrings shouldn't have complex markup." The problem is that a construct that seems superfluous to one person may be absolutely essential to another.

   reStructuredText takes the opposite approach: it provides a rich set of implicit markup constructs (plus a generic extension mechanism for explicit markup), allowing for all kinds of documents. If the set of constructs is too rich for a particular application, the unused constructs can either be removed from the parser (via application-specific overrides) or simply omitted by convention.

3. Why not use indentation for section structure, like StructuredText does? Isn't it more "Pythonic"?

   Guido van Rossum wrote the following in a 2001-06-13 Doc-SIG post:

   I still think that using indentation to indicate sectioning is wrong. If you look at how real books and other print publications are laid out, you'll notice that indentation is used frequently, but mostly at the intra-section level. Indentation can be used to offset lists, tables, quotations, examples, and the like. (The argument that docstrings are different because they are input for a text formatter is wrong: the whole point is that they are also readable without processing.)

   I reject the argument that using indentation is Pythonic: text is not code, and different traditions and conventions hold. People have been presenting text for readability for over 30 centuries. Let's not innovate needlessly.

   See Section Structure via Indentation [25] in Problems With StructuredText [22] for further elaboration.

4. Why use reStructuredText for PEPs? What's wrong with the existing standard?

   The existing standard for PEPs is very limited in terms of general expressibility, and referencing is especially lacking for such a reference-rich document type. PEPs are currently converted into HTML, but the results (mostly monospaced text) are less than attractive, and most of the value-added potential of HTML (especially inline hyperlinks) is untapped.

   Making reStructuredText a standard markup for PEPs will enable much richer expression, including support for section structure, inline markup, graphics, and tables. In several PEPs there are ASCII graphics diagrams, which are all that plaintext documents can support. Since PEPs are made available in HTML form, the ability to include proper diagrams would be immediately useful.

   Current PEP practices allow for reference markers in the form "[1]" in the text, and the footnotes/references themselves are listed in a section toward the end of the document. There is currently no hyperlinking between the reference marker and the footnote/reference itself (it would be possible to add this to pep2html.py, but the "markup" as it stands is ambiguous and mistakes would be inevitable). A PEP with many references (such as this one ;-) requires a lot of flipping back and forth. When revising a PEP, often new references are added or unused references deleted. It is painful to renumber the references, since it has to be done in two places and can have a cascading effect (insert a single new reference 1, and every other reference has to be renumbered; always adding new references to the end is suboptimal). It is easy for references to go out of sync.

   PEPs use references for two purposes: simple URL references and footnotes. reStructuredText differentiates between the two. A PEP might contain references like this:

   Abstract
   
       This PEP proposes adding frungible doodads [1] to the core.
       It extends PEP 9876 [2] via the BCA [3] mechanism.
   
   ...
   
   References and Footnotes
   
       [1] http://www.example.org/
   
       [2] PEP 9876, Let's Hope We Never Get Here
           http://www.python.org/peps/pep-9876.html
   
       [3] "Bogus Complexity Addition"
   

   Reference 1 is a simple URL reference. Reference 2 is a footnote containing text and a URL. Reference 3 is a footnote containing text only. Rewritten using reStructuredText, this PEP could look like this:

   Abstract
   ========
   
   This PEP proposes adding `frungible doodads`_ to the core.  It
   extends PEP 9876 [#pep9876]_ via the BCA [#]_ mechanism.
   
   ...
   
   References & Footnotes
   ======================
   
   .. _frungible doodads: http://www.example.org/
   
   .. [#pep9876] PEP 9876, Let's Hope We Never Get Here
   
   .. [#] "Bogus Complexity Addition"
   

   URLs and footnotes can be defined close to their references if desired, making them easier to read in the source text, and making the PEPs easier to revise. The "References and Footnotes" section can be auto-generated with a document tree transform. Footnotes from throughout the PEP would be gathered and displayed under a standard header. If URL references should likewise be written out explicitly (in citation form), another tree transform could be used.

   URL references can be named ("frungible doodads"), and can be referenced from multiple places in the document without additional definitions. When converted to HTML, references will be replaced with inline hyperlinks (HTML <a> tags). The two footnotes are automatically numbered, so they will always stay in sync. The first footnote also contains an internal reference name, "pep9876", so it's easier to see the connection between reference and footnote in the source text. Named footnotes can be referenced multiple times, maintaining consistent numbering.

   The "#pep9876" footnote could also be written in the form of a citation:

   It extends PEP 9876 [PEP9876]_ ...
   
   .. [PEP9876] PEP 9876, Let's Hope We Never Get Here
   

   Footnotes are numbered, whereas citations use text for their references.

5. Wouldn't it be better to keep the docstring and PEP proposals separate?

   The PEP markup proposal may be removed if it is deemed that there is no need for PEP markup, or it could be made into a separate PEP. If accepted, PEP 1, PEP Purpose and Guidelines [1], and PEP 9, Sample PEP Template [2] will be updated.

   It seems natural to adopt a single consistent markup standard for all uses of structured plaintext in Python, and to propose it all in one place.

6. The existing pep2html.py script converts the existing PEP format to HTML. How will the new-format PEPs be converted to HTML?

   A new version of pep2html.py with integrated reStructuredText parsing has been completed. The Docutils project supports PEPs with a "PEP Reader" component, including all functionality currently in pep2html.py (auto-recognition of PEP & RFC references, email masking, etc.).

7. Who's going to convert the existing PEPs to reStructuredText?

   PEP authors or volunteers may convert existing PEPs if they like, but there is no requirement to do so. The reStructuredText-based PEPs will coexist with the old PEP standard. The pep2html.py mentioned in answer 6 processes both old and new standards.

8. Why use reStructuredText for README and other ancillary files?

   The reasoning given for PEPs in answer 4 above also applies to README and other ancillary files. By adopting a standard markup, these files can be converted to attractive cross-referenced HTML and put up on python.org. Developers of other projects can also take advantage of this facility for their own documentation.

9. Won't the superficial similarity to existing markup conventions cause problems, and result in people writing invalid markup (and not noticing, because the plaintext looks natural)? How forgiving is reStructuredText of "not quite right" markup?

   There will be some mis-steps, as there would be when moving from one programming language to another. As with any language, proficiency grows with experience. Luckily, reStructuredText is a very little language indeed.

   As with any syntax, there is the possibility of syntax errors. It is expected that a user will run the processing system over their input and check the output for correctness.

   In a strict sense, the reStructuredText parser is very unforgiving (as it should be; "In the face of ambiguity, refuse the temptation to guess" [3] applies to parsing markup as well as computer languages). Here's design goal 3 from An Introduction to reStructuredText [19]:

   Unambiguous. The rules for markup must not be open for interpretation. For any given input, there should be one and only one possible output (including error output).

   While unforgiving, at the same time the parser does try to be helpful by producing useful diagnostic output ("system messages"). The parser reports problems, indicating their level of severity (from least to most: debug, info, warning, error, severe). The user or the client software can decide on reporting thresholds; they can ignore low-level problems or cause high-level problems to bring processing to an immediate halt. Problems are reported during the parse as well as included in the output, often with two-way links between the source of the problem and the system message explaining it.

10. Will the docstrings in the Python standard library modules be converted to reStructuredText?

    No. Python's library reference documentation is maintained separately from the source. Docstrings in the Python standard library should not try to duplicate the library reference documentation. The current policy for docstrings in the Python standard library is that they should be no more than concise hints, simple and markup-free (although many do contain ad-hoc implicit markup).

11. I want to write all my strings in Unicode. Will anything break?

    The parser fully supports Unicode. Docutils supports arbitrary input and output encodings.

12. Why does the community need a new structured text design?

    The existing structured text designs are deficient, for the reasons given in "Rationale" above. reStructuredText aims to be a complete markup syntax, within the limitations of the "readable plaintext" medium.

13. What is wrong with existing documentation methodologies?

    What existing methodologies? For Python docstrings, there is no official standard markup format, let alone a documentation methodology akin to JavaDoc. The question of methodology is at a much higher level than syntax (which this PEP addresses). It is potentially much more controversial and difficult to resolve, and is intentionally left out of this discussion.

References & Footnotes

Copyright

This document has been placed in the public domain.

Acknowledgements

Some text is borrowed from PEP 216, Docstring Format [4], by Moshe Zadka.

Special thanks to all members past & present of the Python Doc-SIG [6].