diff --git a/lyse/__init__.py b/lyse/__init__.py
index 6163b8a..cf895e0 100644
--- a/lyse/__init__.py
+++ b/lyse/__init__.py
@@ -76,28 +76,133 @@ class _RoutineStorage(object):
 routine_storage = _RoutineStorage()
 
 
-def data(filepath=None, host='localhost', port=_lyse_port, timeout=5):
+def data(filepath=None, host='localhost', port=_lyse_port, timeout=5, n_sequences=None, filter_kwargs=None):
+    """Get data from the lyse dataframe or a file.
+    
+    This function allows for either extracting information from a run's hdf5
+    file, or retrieving data from lyse's dataframe. If `filepath` is provided
+    then data will be read from that file and returned as a pandas series. If
+    `filepath` is not provided then the dataframe in lyse, or a portion of it,
+    will be returned.
+    
+    Often only part of the lyse dataframe is needed, so the `n_sequences` and
+    `filter_kwargs` arguments provide ways to restrict what parts of the lyse
+    dataframe are returned. The dataframe can be quite large, so only requesting
+    a small part of it can speed up the execution of `lyse.data()` noticeably.
+    Setting `n_sequences` makes this function return only the rows of the lyse
+    dataframe that correspond to the `n_sequences` most recent sequences, where
+    one sequence corresponds to one call to engage in runmanager. Additionally,
+    the `Dataframe.filter()` method can be called on the dataframe before it is
+    transmitted, and the arguments specified in `filter_kwargs` are passed to
+    that method.
+
+    Args:
+        filepath (str, optional): The path to a run's hdf5 file. If a value
+            other than `None` is provided, then this function will return a
+            pandas series containing data associated with the run. In particular
+            it will contain the globals, singleshot results, multishot results,
+            etc. that would appear in the run's row in the Lyse dataframe, but
+            the values will be read from the file rather than extracted from the
+            lyse dataframe. If `filepath` is `None, then this function will
+            instead return a section of the lyse dataframe. Note that if
+            `filepath` is not None, then the other arguments will be ignored.
+            Defaults to `None`.
+        host (str, optional): The address of the computer running lyse. Defaults
+            to `'localhost'`.
+        port (int, optional): The port on which lyse is listening. Defaults to
+            the entry for lyse's port in the labconfig, with a fallback value of
+            42519 if the labconfig has no such entry.
+        timeout (float, optional): The timeout, in seconds, for the
+            communication with lyse. Defaults to 5.
+        n_sequences (int, optional): The maximum number of sequences to include
+            in the returned dataframe where one sequence corresponds to one call
+            to engage in runmanager. The dataframe rows for the most recent
+            `n_sequences` sequences are returned. If the dataframe contains
+            fewer than `n_sequences` sequences, then all rows will be returned.
+            If set to `None`, then all rows are returned. Defaults to `None`.
+        filter_kwargs (dict, optional): A dictionary of keyword arguments to
+            pass to the `Dataframe.filter()` method before the lyse dataframe is
+            returned. For example to call `filter()` with `like='temperature'`,
+            set `filter_kwargs` to `{'like':'temperature'}`. If set to `None`
+            then `Dataframe.filter()` will not be called. See
+            :meth:`pandas:pandas.DataFrame.filter` for more information.
+            Defaults to `None`.
+
+    Raises:
+        ValueError: If `n_sequences` isn't `None` or a nonnegative integer, then
+            a `ValueError` is raised. Note that no `ValueError` is raised if
+            `n_sequences` is greater than the number of sequences available. In
+            that case as all available sequences are returned, i.e. the entire
+            lyse dataframe is returned.
+
+    Returns:
+        :obj:`pandas:pandas.DataFrame` or :obj:`pandas:pandas.Series`: If
+        `filepath` is provided, then a pandas series with the data read from
+        that file is returned. If `filepath` is omitted or set to `None` then
+        the lyse dataframe, or a subset of it, is returned.
+    """    
     if filepath is not None:
         return _get_singleshot(filepath)
     else:
-        df = zmq_get(port, host, 'get dataframe', timeout)
-        try:
-            padding = ('',)*(df.columns.nlevels - 1)
-            try:
-                integer_indexing = _labconfig.getboolean('lyse', 'integer_indexing')
-            except (LabConfig.NoOptionError, LabConfig.NoSectionError):
-                integer_indexing = False
-            if integer_indexing:
-                df.set_index(['sequence_index', 'run number', 'run repeat'], inplace=True, drop=False)
-            else:
-                df.set_index([('sequence',) + padding,('run time',) + padding], inplace=True, drop=False)
-                df.index.names = ['sequence', 'run time']
-        except KeyError:
-            # Empty DataFrame or index column not found, so fall back to RangeIndex instead
-            pass
-        df.sort_index(inplace=True)
+        if n_sequences is not None:
+            if not (type(n_sequences) is int and n_sequences >= 0):
+                msg = """n_sequences must be None or an integer greater than 0 but 
+                    was {n_sequences}.""".format(n_sequences=n_sequences)
+                raise ValueError(dedent(msg))
+        if filter_kwargs is not None:
+            if type(filter_kwargs) is not dict:
+                msg = """filter must be None or a dictionary but was 
+                    {filter_kwargs}.""".format(filter_kwargs=filter_kwargs)
+                raise ValueError(dedent(msg))
+
+        # Allow sending 'get dataframe' (without the enclosing list) if
+        # n_sequences and filter_kwargs aren't provided. This is for backwards
+        # compatability in case the server is running an outdated version of
+        # lyse.
+        if n_sequences is None and filter_kwargs is None:
+            command = 'get dataframe'
+        else:
+            command = ('get dataframe', n_sequences, filter_kwargs)
+        df = zmq_get(port, host, command, timeout)
+        if isinstance(df, str) and df.startswith('error: operation not supported'):
+            # Sending a tuple for command to an outdated lyse servers causes it
+            # to reply with an error message.
+            msg = """The lyse server does not support n_sequences or filter_kwargs.
+                Call this function without providing those arguments to communicate
+                with this server, or upgrade the version of lyse running on the
+                server."""
+            raise ValueError(dedent(msg))
+        # Ensure conversion to multiindex is done, which needs to be done here
+        # if the server is running an outdated version of lyse.
+        _rangeindex_to_multiindex(df, inplace=True)
         return df
-        
+
+def _rangeindex_to_multiindex(df, inplace):
+    if isinstance(df.index, pandas.MultiIndex):
+        # The dataframe has already been converted.
+        return df
+    try:
+        padding = ('',)*(df.columns.nlevels - 1)
+        try:
+            integer_indexing = _labconfig.getboolean('lyse', 'integer_indexing')
+        except (LabConfig.NoOptionError, LabConfig.NoSectionError):
+            integer_indexing = False
+        if integer_indexing:
+            out = df.set_index(['sequence_index', 'run number', 'run repeat'], inplace=inplace, drop=False)
+            # out is None if inplace is True, and is the new dataframe is inplace is False.
+            if not inplace:
+                df = out
+        else:
+            out = df.set_index([('sequence',) + padding,('run time',) + padding], inplace=inplace, drop=False)
+            if not inplace:
+                df = out
+            df.index.names = ['sequence', 'run time']
+    except KeyError:
+        # Empty DataFrame or index column not found, so fall back to RangeIndex instead
+        pass
+    df.sort_index(inplace=True)
+    return df
+
 def globals_diff(run1, run2, group=None):
     return dict_diff(run1.get_globals(group), run2.get_globals(group))
  
diff --git a/lyse/__main__.py b/lyse/__main__.py
index 0d9a943..2d67df0 100644
--- a/lyse/__main__.py
+++ b/lyse/__main__.py
@@ -49,7 +49,7 @@
 from qtutils import inmain_decorator, inmain, UiLoader, DisconnectContextManager
 from qtutils.auto_scroll_to_end import set_auto_scroll_to_end
 import qtutils.icons
-from lyse import LYSE_DIR
+from lyse import LYSE_DIR, _rangeindex_to_multiindex
 
 process_tree = ProcessTree.instance()
 
@@ -156,14 +156,20 @@ def handler(self, request_data):
         logger.info('WebServer request: %s' % str(request_data))
         if request_data == 'hello':
             return 'hello'
+        elif isinstance(request_data, tuple) and request_data[0]=='get dataframe' and len(request_data)==3:
+            _, n_sequences, filter_kwargs = request_data
+            df = self._retrieve_dataframe()
+            df = _rangeindex_to_multiindex(df, inplace=True)
+            # Return only a subset of the dataframe if instructed to do so.
+            if n_sequences is not None:
+                df = self._extract_n_sequences_from_df(df, n_sequences)
+            if filter_kwargs is not None:
+                df = df.filter(**filter_kwargs)
+            return df
         elif request_data == 'get dataframe':
-            # infer_objects() picks fixed datatypes for columns that are compatible with
-            # fixed datatypes, dramatically speeding up pickling. It is called here
-            # rather than when updating the dataframe as calling it during updating may
-            # call it needlessly often, whereas it only needs to be called prior to
-            # sending the dataframe to a client requesting it, as we're doing now.
-            app.filebox.shots_model.infer_objects()
-            return app.filebox.shots_model.dataframe
+            # Ensure backwards compatability with clients using outdated
+            # versions of lyse.
+            return self._retrieve_dataframe()
         elif isinstance(request_data, dict):
             if 'filepath' in request_data:
                 h5_filepath = shared_drive.path_to_local(request_data['filepath'])
@@ -181,6 +187,57 @@ def handler(self, request_data):
         return ("error: operation not supported. Recognised requests are:\n "
                 "'get dataframe'\n 'hello'\n {'filepath': <some_h5_filepath>}")
 
+    @inmain_decorator(wait_for_return=True)
+    def _retrieve_dataframe(self):
+        # infer_objects() picks fixed datatypes for columns that are compatible with
+        # fixed datatypes, dramatically speeding up pickling. It is called here
+        # rather than when updating the dataframe as calling it during updating may
+        # call it needlessly often, whereas it only needs to be called prior to
+        # sending the dataframe to a client requesting it, as we're doing now.
+        app.filebox.shots_model.infer_objects()
+        df = app.filebox.shots_model.dataframe.copy(deep=True)
+        return df
+
+    def _extract_n_sequences_from_df(self, df, n_sequences):
+        # If the dataframe is empty, just return it, otherwise accessing columns
+        # below will raise a KeyError.
+        if df.empty:
+            return df
+
+        # Get a list of all unique sequences, each corresponding to one call to
+        # engage in runmanager. Each sequence may contain multiple runs. The
+        # below creates strings to identify sequences. To be from the same
+        # sequence, two shots have to have the same value for 'sequence' (which
+        # makes sure that the time when engage was called are the same to within
+        # 1 second), 'labscript' (must have been generated from the same
+        # labscript), and 'sequence_index' (a counter which keeps track of how
+        # many times engage has been called and resets to 0 at the start of each
+        # day). Typically just the value for sequence, is enough. However it
+        # only records time down to the second, so if engage() is called twice
+        # quickly then two different sequences can end up with the same value
+        # there.
+        sequences = [str(sequence) for sequence in df['sequence']]
+        labscripts = [str(labscript) for labscript in df['labscript']]
+        sequence_indices = [str(index) for index in df['sequence_index']]
+        # Combine into one string.
+        criteria = zip(sequences, labscripts, sequence_indices)
+        indentity_strings = [seq + script + ind for seq, script, ind in criteria]
+
+        # Find the distinct values, maintaining their ordering.
+        unique_identities = np.intersect1d(indentity_strings, indentity_strings)
+
+        # Slice the DataFrame so that only the last n_sequences sequences
+        # remain. Note that slicing unique_identities just returns all of its
+        # entries if n_sequences is greater than its length; it doesn't raise an
+        # error.
+        if n_sequences == 0:
+            identities_included = []
+        else:
+            identities_included = unique_identities[-n_sequences:]
+        df_subset = df[[id in identities_included for id in indentity_strings]]
+
+        return df_subset
+
 
 class LyseMainWindow(QtWidgets.QMainWindow):
     # A signal to show that the window is shown and painted.