Fork me on GitHub

Features and Caveats

Jedi obviously supports autocompletion. It’s also possible to get it working in (your REPL (IPython, etc.)).

Static analysis is also possible by using the command jedi.names.

The Jedi Linter is currently in an alpha version and can be tested by calling python -m jedi linter.

Jedi would in theory support refactoring, but we have never publicized it, because it’s not production ready. If you’re interested in helping out here, let me know. With the latest parser changes, it should be very easy to actually make it work.

General Features

  • python 2.6+ and 3.2+ support
  • ignores syntax errors and wrong indentation
  • can deal with complex module / function / class structures
  • virtualenv support
  • can infer function arguments from sphinx, epydoc and basic numpydoc docstrings (type hinting)

Supported Python Features

Jedi supports many of the widely used Python features:

  • builtins
  • multiple returns or yields
  • tuple assignments / array indexing / dictionary indexing
  • with-statement / exception handling
  • *args / **kwargs
  • decorators / lambdas / closures
  • generators / iterators
  • some descriptors: property / staticmethod / classmethod
  • some magic methods: __call__, __iter__, __next__, __get__, __getitem__, __init__
  • list.append(), set.add(), list.extend(), etc.
  • (nested) list comprehensions / ternary expressions
  • relative imports
  • getattr() / __getattr__ / __getattribute__
  • function annotations (py3k feature, are ignored right now, but being parsed. I don’t know what to do with them.)
  • class decorators (py3k feature, are being ignored too, until I find a use case, that doesn’t work with Jedi)
  • simple/usual sys.path modifications
  • isinstance checks for if/while/assert
  • namespace packages (includes pkgutil and pkg_resources namespaces)
  • Django / Flask / Buildout support

Unsupported Features

Not yet implemented:

  • manipulations of instances outside the instance variables without using methods

Will probably never be implemented:

  • metaclasses (how could an auto-completion ever support this)
  • setattr(), __import__()
  • writing to some dicts: globals(), locals(), object.__dict__
  • evaluating if / while / del


Malformed Syntax

Syntax errors and other strange stuff may lead to undefined behaviour of the completion. Jedi is NOT a Python compiler, that tries to correct you. It is a tool that wants to help you. But YOU have to know Python, not Jedi.

Legacy Python 2 Features

This framework should work for both Python 2/3. However, some things were just not as pythonic in Python 2 as things should be. To keep things simple, some older Python 2 features have been left out:

  • Classes: Always Python 3 like, therefore all classes inherit from object.
  • Generators: No next() method. The __next__() method is used instead.

Slow Performance

Importing numpy can be quite slow sometimes, as well as loading the builtins the first time. If you want to speed things up, you could write import hooks in Jedi, which preload stuff. However, once loaded, this is not a problem anymore. The same is true for huge modules like PySide, wx, etc.


Security is an important issue for Jedi. Therefore no Python code is executed. As long as you write pure python, everything is evaluated statically. But: If you use builtin modules (c_builtin) there is no other option than to execute those modules. However: Execute isn’t that critical (as e.g. in pythoncomplete, which used to execute every import!), because it means one import and no more. So basically the only dangerous thing is using the import itself. If your c_builtin uses some strange initializations, it might be dangerous. But if it does you’re screwed anyways, because eventualy you’re going to execute your code, which executes the import.


Here are some tips on how to use Jedi efficiently.

Type Hinting

If Jedi cannot detect the type of a function argument correctly (due to the dynamic nature of Python), you can help it by hinting the type using one of the following docstring syntax styles:

Sphinx style

def myfunction(node, foo):
    """Do something with a ``node``.

    :type node: ProgramNode
    :param str foo: foo parameter description

    node.| # complete here


def myfunction(node):
    """Do something with a ``node``.

    @type node: ProgramNode

    node.| # complete here


In order to support the numpydoc format, you need to install the numpydoc package.

def foo(var1, var2, long_var_name='hi'):
    r"""A one-line summary that does not use variable names or the
    function name.


    var1 : array_like
        Array_like means all those objects -- lists, nested lists,
        etc. -- that can be converted to an array. We can also
        refer to variables like `var1`.
    var2 : int
        The type above can either refer to an actual Python type
        (e.g. ``int``), or describe the type of the variable in more
        detail, e.g. ``(N,) ndarray`` or ``array_like``.
    long_variable_name : {'hi', 'ho'}, optional
        Choices in brackets, default first when optional.


    var2.| # complete here

A little history

The Star Wars Jedi are awesome. My Jedi software tries to imitate a little bit of the precognition the Jedi have. There’s even an awesome scene of Monty Python Jedis :-).

But actually the name hasn’t so much to do with Star Wars. It’s part of my second name.

After I explained Guido van Rossum, how some parts of my auto-completion work, he said (we drank a beer or two):

“Oh, that worries me...”

When it’s finished, I hope he’ll like it :-)

I actually started Jedi, because there were no good solutions available for VIM. Most auto-completions just didn’t work well. The only good solution was PyCharm. But I like my good old VIM. Rope was never really intended to be an auto-completion (and also I really hate project folders for my Python scripts). It’s more of a refactoring suite. So I decided to do my own version of a completion, which would execute non-dangerous code. But I soon realized, that this wouldn’t work. So I built an extremely recursive thing which understands many of Python’s key features.

By the way, I really tried to program it as understandable as possible. But I think understanding it might need quite some time, because of its recursive nature.