Consistent and faster stories fingerprinting

changelog/4612.bugfix.md

@@ -0,0 +1 @@
+Fixed the bug that OR statements in stories would break the check whether a model needs to be retrained

changelog/7955.improvement.md

@@ -0,0 +1 @@
+Drastically improved finger printing time for large story graphs

data/test_yaml_stories/story_with_two_equal_or_statements.yml

@@ -0,0 +1,25 @@
+stories:
+- story: story_with_two_equal_or_statements
+  steps:
+  - intent: simple
+  - action: utter_default
+  - or:
+    - intent: affirm
+    - intent: thank_you
+  - action: utter_greet
+  - action: utter_default
+  - or:
+    - intent: affirm
+    - intent: thank_you
+  - action: utter_goodbye
+# same name again, same events, same intents, but with entities
+- story: story_with_two_equal_or_statements
+  steps:
+  - intent: simple
+  - action: utter_default
+  - or:
+    - intent: affirm
+      entities:
+        - name: peter
+    - intent: thank_you
+  - action: utter_greet

rasa/shared/core/training_data/loading.py

@@ -54,7 +54,7 @@ def _guess_reader(
 
 
 async def load_data_from_resource(
-    resource: Union[Text, Path],
+    resource: Union[Text],
     domain: Domain,
     template_variables: Optional[Dict] = None,
     use_e2e: bool = False,

rasa/shared/core/training_data/story_reader/story_step_builder.py

@@ -91,9 +91,8 @@ def add_user_messages(
             # user can use the express the same thing
             # we need to copy the blocks and create one
             # copy for each possible message
-            prefix = GENERATED_CHECKPOINT_PREFIX + "OR_"
-            generated_checkpoint = rasa.shared.core.training_data.structures.generate_id(
-                prefix, GENERATED_HASH_LENGTH
+            generated_checkpoint = self._generate_checkpoint_name_for_or_statement(
+                messages
             )
             updated_steps = []
             for t in self.current_steps:
@@ -143,3 +142,35 @@ def _next_story_steps(self) -> List[StoryStep]:
             )
         ]
         return current_turns
+
+    def _generate_checkpoint_name_for_or_statement(
+        self, messages: List[UserUttered]
+    ) -> str:
+        """Generates a unique checkpoint name for an or statement.
+
+        The name is based on the current story/rule name,
+        the current place in the story since the last checkpoint or start,
+        the name of the starting checkpoints,
+        and the involved intents/e2e messages.
+        """
+        messages_texts_or_intents = sorted([str(m) for m in messages])
+        start_checkpoint_names = sorted(
+            list({chk.name for s in self.current_steps for chk in s.start_checkpoints})
+        )
+        events = [str(e) for s in self.current_steps for e in s.events]
+        # name: to identify the current story or rule
+        # events: to identify what has happened so far
+        #         within the current story/rule
+        # start checkpoint names: to identify the section
+        #                         within the current story/rule when there are
+        #                         multiple internal checkpoints
+        # messages texts or intents: identifying the members of the or statement
+        unique_id = (
+            f"{self.name}_<>_{'@@@'.join(events)}"
+            f"_<>_{'@@@'.join(start_checkpoint_names)}"
+            f"_<>_{'@@@'.join(messages_texts_or_intents)}"
+        )
+        hashed_id = rasa.shared.utils.io.get_text_hash(unique_id)[
+            :GENERATED_HASH_LENGTH
+        ]
+        return f"{GENERATED_CHECKPOINT_PREFIX}OR_{hashed_id}"

rasa/shared/core/training_data/structures.py

@@ -411,8 +411,8 @@ def fingerprint(self) -> Text:
             YAMLStoryWriter,
         )
 
-        self_as_string = YAMLStoryWriter().dumps(self.story_steps)
-        return rasa.shared.utils.io.get_text_hash(self_as_string)
+        stories_as_yaml = YAMLStoryWriter().stories_to_yaml(self.story_steps)
+        return rasa.shared.utils.io.deep_container_fingerprint(stories_as_yaml)
 
     def ordered_steps(self) -> List[StoryStep]:
         """Returns the story steps ordered by topological order of the DAG."""

tests/shared/core/training_data/test_graph.py

@@ -1,4 +1,6 @@
 from rasa.shared.core.training_data.structures import StoryGraph
+import rasa.shared.core.training_data.loading
+from rasa.shared.core.domain import Domain
 
 
 def check_graph_is_sorted(g, sorted_nodes, removed_edges):
@@ -45,3 +47,43 @@ def test_node_ordering_with_cycle():
 
 def test_is_empty():
     assert StoryGraph([]).is_empty()
+
+
+async def test_consistent_fingerprints():
+    stories_path = "data/test_yaml_stories/stories.yml"
+    domain_path = "data/test_domains/default_with_slots.yml"
+    domain = Domain.load(domain_path)
+    story_steps = await rasa.shared.core.training_data.loading.load_data_from_resource(
+        stories_path, domain
+    )
+    story_graph = StoryGraph(story_steps)
+
+    # read again
+    story_steps_2 = await rasa.shared.core.training_data.loading.load_data_from_resource(
+        stories_path, domain
+    )
+    story_graph_2 = StoryGraph(story_steps_2)
+
+    fingerprint = story_graph.fingerprint()
+    fingerprint_2 = story_graph_2.fingerprint()
+
+    assert fingerprint == fingerprint_2
+
+
+async def test_unique_checkpoint_names():
+    stories_path = "data/test_yaml_stories/story_with_two_equal_or_statements.yml"
+    domain_path = "data/test_domains/default_with_slots.yml"
+    domain = Domain.load(domain_path)
+    story_steps = await rasa.shared.core.training_data.loading.load_data_from_resource(
+        stories_path, domain
+    )
+    start_checkpoint_names = {
+        chk.name for s in story_steps for chk in s.start_checkpoints
+    }
+
+    # first story:
+    # START_CHECKPOINT, GENR_OR_XXXXX for first OR, GENR_OR_YYYYY for second OR
+
+    # additional in second story:
+    # GENR_OR_ZZZZZ as entities are different from first OR in first story
+    assert len(start_checkpoint_names) == 4