Roles and Directives


Create a citation to a bibliographic entry. For example:

See :cite:`1987:nelson` for an introduction to non-standard analysis.

which would be equivalent to the following LaTeX code:

See \cite{1987:nelson} for an introduction to non-standard analysis.

Multiple comma-separated keys can be specified at once:

See :cite:`1987:nelson,2001:schechter`.


Due to a docutils implementation detail, Sphinx’s LaTeX backend will not actually generate \cite commands. Instead, all references, including citation references, are managed using \hyperref and \label commands. See

.. bibliography:: refs.bib [...]

Create bibliography for all cited references. The all flag forces all references to be included (equivalent to \nocite{*} in LaTeX). The notcited flag causes all references that were not cited to be included. The cited flag is recognized as well but is entirely optional. For example:

.. rubric:: References

.. bibliography:: refs.bib

which would be roughly equivalent to the following LaTeX code:

  \newblock {\em Radically Elementary Probability Theory}.
  \newblock Princeton University Press, 1987.

Note that, unlike LaTeX, the bibliography directive does not generate a default section title.


Sphinx may not be able to create an entry for cite keys when your bibliography directive resides in a different document; see Unresolved Citations Across Documents for more information and workarounds.

You can also pick a bibliography style, using the style option. The alpha style is the default. Other supported styles are plain, unsrt, and unsrtalpha.

.. bibliography:: refs.bib
   :style: unsrt


Sphinx will attempt to resolve references to the bibliography across all documents, so you must take care that no citation key is included more than once.

You can also set the encoding of the bibliography files, using the encoding option.

.. bibliography:: refs.bib
   :encoding: latex+latin

Note that, usually, you want to prepend your encoding with latex+, in order to convert LaTeX control characters to unicode characters (for instance, to convert \'e into é). The latex codec is invoked by default, for your convenience. Be sure to write \% when you intend to format a percent sign.

Advanced Features

Bullet Lists and Enumerated Lists

New in version 0.2.4.

You can change the type of list used for rendering the bibliography. By default, a paragraph of standard citations is generated. However, instead, you can also generate a bullet list, or an enumerated list.

.. bibliography:: refs1.bib
   :list: bullet

.. bibliography:: refs2.bib
   :list: enumerated

Note that citations to these types of bibliography lists will not be resolved.

For enumerated lists, you can also specify the type (default is arabic), and the start of the sequence (default is 1).

.. bibliography:: refs2.bib
   :list: enumerated
   :enumtype: upperroman
   :start: 3

The enumtype can be any of arabic (1, 2, 3, ...), loweralpha (a, b, c, ...), upperalpha (A, B, C, ...), lowerroman (i, ii, iii, ...), or upperroman (I, II, III, ...).

The start can be any positive integer (1, 2, 3, ...) or continue if you wish the enumeration to continue from the last bibliography directive. This is helpful if you split up your bibliography but still want to enumerate the entries continuously.

Label Prefixing

New in version 0.2.5.

If you have multiple bibliographies, and experience duplicate labels, use the labelprefix option.

.. rubric:: References

.. bibliography:: refs.bib
   :labelprefix: A

.. rubric:: Further reading

.. bibliography:: refs.bib
   :labelprefix: B


New in version 0.2.7.

Whilst the cited, all, and notcited options will cover many use cases, sometimes more advanced selection of bibliographic entries is desired. For this purpose, you can use the filter option:

.. bibliography:: refs.bib
   :list: bullet
   :filter: author % "Einstein"

The string specified in the filter option must be a valid Python expression.


The expression is parsed using ast.parse() and then evaluated using an ast.NodeVisitor, so it should be reasonably safe against malicious code.

The filter expression supports:

  • The boolean operators and, or.

  • The unary operator not.

  • The comparison operators ==, <=, <, >=, and >.

  • Regular expression matching using the % operator, where the left hand side is the string to be matched, and the right hand side is the regular expression. Matching is case insensitive. For example:

    .. bibliography:: refs.bib
       :list: bullet
       :filter: title % "relativity"

    would include all entries that have the word “relativity” in the title.


    The implementation uses

  • Single and double quoted strings, such as 'hello' or "world".

  • Set literals, such has {"hello", "world"}, as well as the set operators &, |, in, and not in.

    New in version 0.3.0.

  • Various identifiers, such as:

    • type is the entry type, as a lower case string (i.e. "inproceedings").

    • key is the entry key, as a lower case string (this is because keys are considered case insensitive).

    • cited evaluates to True if the entry was cited in the document, and to False otherwise.

    • docname evaluates to the name of the current document.

      New in version 0.3.0.

    • docnames evaluates to a set of names from which the entry is cited.

      New in version 0.3.0.

    • True and False.

    • author is the entry string of authors in standard format (last, first), separated by “and”.

    • editor is similar to author but for editors.

    • Any other (lower case) identifier evaluates to a string containing the value of the correspondingly named field, such as title, publisher, year, and so on. If the item is missing in the entry then it evaluates to the empty string. Here is an example of how one would typically write an expression to filter on an optional field:

      .. bibliography:: refs.bib
         :list: bullet
         :filter: cited and year and (year <= "2003")

      which would include all cited entries that have a year that is less or equal than 2003; any entries that do not specify a year would be omitted.

Local Bibliographies

To create a bibliography that includes only citations that were cited in the current document, use the following filter:

.. bibliography:: refs.bib
   :filter: docname in docnames

More generally, you can create bibliographies for citations that were cited from specific documents only:

.. bibliography:: refs.bib
   :filter: {"doc1", "doc2"} & docnames

This bibliography will include all citations that were cited from doc1.rst or doc2.rst. Another hypothetical example:

.. bibliography:: refs.bib
   :filter: cited and ({"doc1", "doc2"} >= docnames)

This bibliography will include all citations that were cited in doc1.rst or doc2.rst, but nowhere else.

Custom Formatting, Sorting, and Labelling

pybtex provides a very powerful way to create and register new styles, using setuptools entry points, as documented here:

Simply add the following code to your

from import Style as UnsrtStyle
from import toplevel # ... and anything else needed
from pybtex.plugin import register_plugin

class MyStyle(UnsrtStyle):

    def format_XXX(self, e):
        template = toplevel [
            # etc.
        return template.format_data(e)

register_plugin('', 'mystyle', MyStyle)

Now mystyle will be available to you as a formatting style:

.. bibliography:: refs.bib
   :style: mystyle

The formatting code uses a very intuitive template engine. The source code for unsrt provides many great examples:

The above example only demonstrates a custom formatting style plugin. It is also possible to register custom author/editor naming plugins (using the group) labelling plugins (using the group), and sorting plugins (using the group).

An minimal example is available here:

Known Issues and Workarounds


To use the bibtex extension with Tinkerer, be sure to specify the bibtex extension first in your file:

extensions = ['sphinxcontrib.bibtex', '', 'tinkerer.ext.disqus']

Encoding: Percent Signs

When using the LaTeX codec (which is by default), be sure to write \% for percent signs at all times (unless your file contains a genuine comment), otherwise the bibtex lexer will ignore the remainder of the line.

If you don’t want any LaTeX symbols to be reinterpreted as unicode, use the option :encoding: utf (without the latex+ prefix).

Unresolved Citations Across Documents

If you cite something that has its bibliography in another document, then, at the moment, the extension may, or may not, realise that it has to add this citation. There are a few ways to work around this problem:

  • Use the option :all: in the bibliography directive (which will simply cause all entries to be included).
  • Ensure that the bibliography directive is processed after all cites. Sphinx appears to process files in an alphabetical manner. For instance, in case you have only one file containing a bibliography directive, simply name that file zreferences.rst.

Hopefully, this limitation can be lifted in a future release.

Duplicate Labels When Using :style: plain

With :style: plain, labels are numerical, restarting at [1] for each bibliography directive. Consequently, when inserting multiple bibliography directives with :style: plain, you are bound to get duplicate labels for entries. There are a few ways to work around this problem:

  • Use a single bibliography directive for all your references.
  • Use the labelprefix option, as documented above.
  • Use a style that has non-numerical labelling, such as :style: alpha.

Mismatch Between Output of HTML and LaTeX Backends

Sphinx’s LaTeX writer currently collects all citations together, and puts them on a separate page, with a separate title, whereas the html writer puts citations at the location where they are defined. This issue will occur also if you use regular citations in Sphinx: it has nothing to do with sphinxcontrib-bibtex per se.

To get a closer match between the two outputs, you can tell Sphinx to generate a rubric title only for html:

.. only:: html

   .. rubric:: References

.. bibliography:: refs.bib

This code could be placed in your zreferences.rst.

The current aim is to fix Sphinx’s LaTeX writer to match the html output more closely. The issue is tracked here: