diff --git a/jedi/inference/gradual/typing.py b/jedi/inference/gradual/typing.py
index 64e1e43c..2d7d18ef 100644
--- a/jedi/inference/gradual/typing.py
+++ b/jedi/inference/gradual/typing.py
@@ -389,7 +389,13 @@ class TypedDict(LazyValueWrapper):
         return ValueName(self, self.tree_node.name)
 
     def py__simple_getitem__(self, index):
-        return ValueSet({self.inference_state.builtins_module})
+        if isinstance(index, str):
+            return ValueSet.from_sets(
+                name.infer()
+                for filter in self._definition_class.get_filters(is_instance=True)
+                for name in filter.get(index)
+            )
+        return NO_VALUES
 
     def get_key_values(self):
         from jedi.inference.compiled import create_simple_object
diff --git a/jedi/inference/value/instance.py b/jedi/inference/value/instance.py
index 2eca9f9d..b884b14d 100644
--- a/jedi/inference/value/instance.py
+++ b/jedi/inference/value/instance.py
@@ -547,10 +547,10 @@ class InstanceClassFilter(AbstractFilter):
         self._class_filter = class_filter
 
     def get(self, name):
-        return self._convert(self._class_filter.get(name, from_instance=True))
+        return self._convert(self._class_filter.get(name))
 
     def values(self):
-        return self._convert(self._class_filter.values(from_instance=True))
+        return self._convert(self._class_filter.values())
 
     def _convert(self, names):
         return [
@@ -586,7 +586,7 @@ class SelfAttributeFilter(ClassFilter):
             if trailer.type == 'trailer' \
                     and len(trailer.parent.children) == 2 \
                     and trailer.children[0] == '.':
-                if name.is_definition() and self._access_possible(name, from_instance=True):
+                if name.is_definition() and self._access_possible(name):
                     # TODO filter non-self assignments instead of this bad
                     #      filter.
                     if self._is_in_right_scope(trailer.parent.children[0], name):
diff --git a/jedi/inference/value/klass.py b/jedi/inference/value/klass.py
index 764644d6..41c669d7 100644
--- a/jedi/inference/value/klass.py
+++ b/jedi/inference/value/klass.py
@@ -104,12 +104,12 @@ class ClassFilter(ParserTreeFilter):
             node = get_cached_parent_scope(self._used_names, node)
         return False
 
-    def _access_possible(self, name, from_instance=False):
+    def _access_possible(self, name):
         # Filter for ClassVar variables
         # TODO this is not properly done, yet. It just checks for the string
         # ClassVar in the annotation, which can be quite imprecise. If we
         # wanted to do this correct, we would have to infer the ClassVar.
-        if not from_instance:
+        if not self._is_instance:
             expr_stmt = name.get_definition()
             if expr_stmt is not None and expr_stmt.type == 'expr_stmt':
                 annassign = expr_stmt.children[1]
@@ -122,9 +122,9 @@ class ClassFilter(ParserTreeFilter):
         return not name.value.startswith('__') or name.value.endswith('__') \
             or self._equals_origin_scope()
 
-    def _filter(self, names, from_instance=False):
+    def _filter(self, names):
         names = super(ClassFilter, self)._filter(names)
-        return [name for name in names if self._access_possible(name, from_instance)]
+        return [name for name in names if self._access_possible(name)]
 
 
 class ClassMixin(object):
@@ -270,6 +270,7 @@ class ClassValue(use_metaclass(CachedMetaClass, ClassMixin, FunctionAndClassBase
             self.inference_state.builtins_module.py__getattribute__('object')
         )]
 
+    @inference_state_method_cache(default=False)
     def is_typeddict(self):
         # TODO Do a proper mro resolution. Currently we are just listing
         # classes. However, it's a complicated algorithm.
diff --git a/test/completion/pep0484_typing.py b/test/completion/pep0484_typing.py
index 8dc03d96..eef0bf13 100644
--- a/test/completion/pep0484_typing.py
+++ b/test/completion/pep0484_typing.py
@@ -508,7 +508,7 @@ dynamic_annotation('')
 
 class Foo(typing.TypedDict):
     foo: str
-    bar: List[int]
+    bar: typing.List[float]
     foo
     #! ['foo: str']
     foo
@@ -519,7 +519,7 @@ class Foo(typing.TypedDict):
 d: Foo
 #? str()
 d['foo']
-#? str()
+#? float()
 d['bar'][0]
 #?
 d['baz']