diff --git a/python/lammps.py b/python/lammps.py
index 1b0b34691c..a39d96c69d 100644
--- a/python/lammps.py
+++ b/python/lammps.py
@@ -481,6 +481,40 @@ class lammps(object):
           return np.intc
 
         def extract_atom(self, name, dtype=LAMMPS_AUTODETECT, nelem=LAMMPS_AUTODETECT, dim=LAMMPS_AUTODETECT):
+          """Retrieve per-atom properties from LAMMPS as NumPy arrays
+
+          This is a wrapper around the :cpp:func:`lammps_extract_atom`
+          function of the C-library interface. Its documentation includes a
+          list of the supported keywords and their data types.
+          Since Python needs to know the data type to be able to interpret
+          the result, by default, this function will try to auto-detect the datatype
+          by asking the library. You can also force a specific datatype.  For
+          that purpose the :py:mod:`lammps` module contains the constants
+          ``LAMMPS_INT``, ``LAMMPS_DOUBLE``, ``LAMMPS_BIGINT``,
+          ``LAMMPS_TAGINT``, and ``LAMMPS_STRING``.
+          This function returns ``None`` if either the keyword is not
+          recognized, or an invalid data type constant is used.
+
+          .. note::
+
+             While the returned arrays of per-atom data are dimensioned
+             for the range [0:nmax] - as is the underlying storage -
+             the data is usually only valid for the range of [0:nlocal],
+             unless the property of interest is also updated for ghost
+             atoms.  In some cases, this depends on a LAMMPS setting, see
+             for example :doc:`comm_modify vel yes <comm_modify>`.
+
+          :param name: name of the property
+          :type name:  string
+          :param dtype: type of the returned data
+          :type dtype:  int, optional
+          :param nelem: number of elements in array
+          :type nelem:  int, optional
+          :param dim: dimension of each element
+          :type dim:  int, optional
+          :return: requested data as NumPy array with direct access to C data
+          :rtype: numpy.array
+          """
           if dtype == LAMMPS_AUTODETECT:
             dtype = self.lmp.extract_atom_datatype(name)
 
@@ -491,7 +525,11 @@ class lammps(object):
               nelem = self.lmp.extract_global("nlocal")
           if dim == LAMMPS_AUTODETECT:
             if dtype in (LAMMPS_INT2D, LAMMPS_DOUBLE2D, LAMMPS_TAGINT2D):
-              dim = 2
+              # TODO add other fields
+              if name in ("x", "v", "f", "angmom", "torque", "csforce", "vforce"):
+                dim = 3
+              else:
+                dim = 2
             else:
               dim = 1
 
diff --git a/unittest/python/python-numpy.py b/unittest/python/python-numpy.py
index 3e223af877..3c8ff9f512 100644
--- a/unittest/python/python-numpy.py
+++ b/unittest/python/python-numpy.py
@@ -128,8 +128,8 @@ class PythonNumpy(unittest.TestCase):
         ntypes = self.lmp.extract_global("ntypes")
         self.assertEqual(ntypes, 1)
 
-        x = self.lmp.numpy.extract_atom("x", dim=3)
-        v = self.lmp.numpy.extract_atom("v", dim=3)
+        x = self.lmp.numpy.extract_atom("x")
+        v = self.lmp.numpy.extract_atom("v")
         self.assertEqual(len(x), 2)
         self.assertTrue((x[0] == (1.0, 1.0, 1.0)).all())
         self.assertTrue((x[1] == (1.0, 1.0, 1.5)).all())