rasa-pro 3.8.16__py3-none-any.whl

rasa/core/policies/rule_policy.py

@@ -0,0 +1,1276 @@
+from __future__ import annotations
+import copy
+import functools
+import logging
+import structlog
+from typing import Any, List, DefaultDict, Dict, Text, Optional, Set, Tuple, cast
+
+from tqdm import tqdm
+import numpy as np
+import json
+from collections import defaultdict
+
+from rasa.engine.graph import ExecutionContext
+from rasa.engine.recipes.default_recipe import DefaultV1Recipe
+from rasa.engine.storage.resource import Resource
+from rasa.engine.storage.storage import ModelStorage
+from rasa.shared.constants import DOCS_URL_RULES
+from rasa.shared.exceptions import RasaException
+import rasa.shared.utils.io
+from rasa.shared.core.events import LoopInterrupted, UserUttered, ActionExecuted
+from rasa.core.featurizers.tracker_featurizers import TrackerFeaturizer
+from rasa.core.policies.memoization import MemoizationPolicy
+from rasa.core.policies.policy import SupportedData, PolicyPrediction
+from rasa.shared.core.trackers import (
+    DialogueStateTracker,
+    get_active_loop_name,
+    is_prev_action_listen_in_state,
+)
+from rasa.shared.core.generator import TrackerWithCachedStates
+from rasa.core.constants import (
+    DEFAULT_CORE_FALLBACK_THRESHOLD,
+    RULE_POLICY_PRIORITY,
+    POLICY_PRIORITY,
+    POLICY_MAX_HISTORY,
+)
+from rasa.shared.core.constants import (
+    USER_INTENT_RESTART,
+    USER_INTENT_BACK,
+    USER_INTENT_SESSION_START,
+    ACTION_LISTEN_NAME,
+    ACTION_RESTART_NAME,
+    ACTION_SESSION_START_NAME,
+    ACTION_DEFAULT_FALLBACK_NAME,
+    ACTION_BACK_NAME,
+    RULE_SNIPPET_ACTION_NAME,
+    SHOULD_NOT_BE_SET,
+    PREVIOUS_ACTION,
+    LOOP_NAME,
+    SLOTS,
+    ACTIVE_LOOP,
+    RULE_ONLY_SLOTS,
+    RULE_ONLY_LOOPS,
+)
+from rasa.shared.core.domain import InvalidDomain, State, Domain
+from rasa.shared.nlu.constants import ACTION_NAME, INTENT_NAME_KEY
+import rasa.core.test
+from rasa.core.training.training import create_action_fingerprints, ActionFingerprint
+
+logger = logging.getLogger(__name__)
+structlogger = structlog.get_logger()
+
+
+# These are Rasa Open Source default actions and overrule everything at any time.
+DEFAULT_ACTION_MAPPINGS = {
+    USER_INTENT_RESTART: ACTION_RESTART_NAME,
+    USER_INTENT_BACK: ACTION_BACK_NAME,
+    USER_INTENT_SESSION_START: ACTION_SESSION_START_NAME,
+}
+
+RULES = "rules"
+RULES_FOR_LOOP_UNHAPPY_PATH = "rules_for_loop_unhappy_path"
+RULES_NOT_IN_STORIES = "rules_not_in_stories"
+
+LOOP_WAS_INTERRUPTED = "loop_was_interrupted"
+DO_NOT_PREDICT_LOOP_ACTION = "do_not_predict_loop_action"
+
+DEFAULT_RULES = "predicting default action with intent "
+LOOP_RULES = "handling active loops and forms - "
+LOOP_RULES_SEPARATOR = " - "
+
+
+class InvalidRule(RasaException):
+    """Exception that can be raised when rules are not valid."""
+
+    def __init__(self, message: Text) -> None:
+        super().__init__()
+        self.message = message
+
+    def __str__(self) -> Text:
+        return self.message + (
+            f"\nYou can find more information about the usage of "
+            f"rules at {DOCS_URL_RULES}. "
+        )
+
+
+@DefaultV1Recipe.register(
+    DefaultV1Recipe.ComponentType.POLICY_WITHOUT_END_TO_END_SUPPORT, is_trainable=True
+)
+class RulePolicy(MemoizationPolicy):
+    """Policy which handles all the rules."""
+
+    # rules use explicit json strings
+    ENABLE_FEATURE_STRING_COMPRESSION = False
+
+    # number of user inputs that is allowed in case rules are restricted
+    ALLOWED_NUMBER_OF_USER_INPUTS = 1
+
+    @staticmethod
+    def supported_data() -> SupportedData:
+        """The type of data supported by this policy.
+
+        Returns:
+            The data type supported by this policy (ML and rule data).
+        """
+        return SupportedData.ML_AND_RULE_DATA
+
+    @staticmethod
+    def get_default_config() -> Dict[Text, Any]:
+        """Returns the default config (see parent class for full docstring)."""
+        return {
+            # Priority of the policy which is used if multiple policies predict
+            # actions with the same confidence.
+            POLICY_PRIORITY: RULE_POLICY_PRIORITY,
+            # Confidence of the prediction if no rule matched and de-facto
+            # threshold for a core fallback.
+            "core_fallback_threshold": DEFAULT_CORE_FALLBACK_THRESHOLD,
+            # Name of the action which should be predicted if no rule matched.
+            "core_fallback_action_name": ACTION_DEFAULT_FALLBACK_NAME,
+            # If `True` `core_fallback_action_name` is predicted in case no rule
+            # matched.
+            "enable_fallback_prediction": True,
+            # If `True` rules are restricted to contain a maximum of 1
+            # user message. This is used to avoid that users build a state machine
+            # using the rules.
+            "restrict_rules": True,
+            # Whether to check for contradictions between rules and stories
+            "check_for_contradictions": True,
+            # the policy will use the confidence of NLU on the latest
+            # user message to set the confidence of the action
+            "use_nlu_confidence_as_score": False,
+        }
+
+    def __init__(
+        self,
+        config: Dict[Text, Any],
+        model_storage: ModelStorage,
+        resource: Resource,
+        execution_context: ExecutionContext,
+        featurizer: Optional[TrackerFeaturizer] = None,
+        lookup: Optional[Dict] = None,
+    ) -> None:
+        """Initializes the policy."""
+        # max history is set to `None` in order to capture any lengths of rule stories
+        config[POLICY_MAX_HISTORY] = None
+
+        super().__init__(
+            config, model_storage, resource, execution_context, featurizer, lookup
+        )
+
+        self._fallback_action_name = config["core_fallback_action_name"]
+        self._enable_fallback_prediction = config["enable_fallback_prediction"]
+        self._check_for_contradictions = config["check_for_contradictions"]
+
+        self._rules_sources: DefaultDict[Text, List[Tuple[Text, Text]]] = defaultdict(
+            list
+        )
+
+    @classmethod
+    def raise_if_incompatible_with_domain(
+        cls, config: Dict[Text, Any], domain: Domain
+    ) -> None:
+        """Checks whether the domains action names match the configured fallback.
+
+        Args:
+            config: configuration of a `RulePolicy`
+            domain: a domain
+        Raises:
+            `InvalidDomain` if this policy is incompatible with the domain
+        """
+        fallback_action_name = config.get("core_fallback_action_name", None)
+        if (
+            fallback_action_name
+            and fallback_action_name not in domain.action_names_or_texts
+        ):
+            raise InvalidDomain(
+                f"The fallback action '{fallback_action_name}' which was "
+                f"configured for the {RulePolicy.__name__} must be "
+                f"present in the domain."
+            )
+
+    @staticmethod
+    def _is_rule_snippet_state(state: State) -> bool:
+        prev_action_name = state.get(PREVIOUS_ACTION, {}).get(ACTION_NAME)
+        return prev_action_name == RULE_SNIPPET_ACTION_NAME
+
+    def _create_feature_key(self, states: List[State]) -> Optional[Text]:
+        new_states: List[State] = []
+        for state in reversed(states):
+            if self._is_rule_snippet_state(state):
+                # remove all states before RULE_SNIPPET_ACTION_NAME
+                break
+            new_states.insert(0, state)
+
+        if not new_states:
+            return None
+
+        # we sort keys to make sure that the same states
+        # represented as dictionaries have the same json strings
+        return json.dumps(new_states, sort_keys=True)
+
+    @staticmethod
+    def _states_for_unhappy_loop_predictions(states: List[State]) -> List[State]:
+        """Modifies the states to create feature keys for loop unhappy path conditions.
+
+        Args:
+            states: a representation of a tracker
+                as a list of dictionaries containing features
+
+        Returns:
+            modified states
+        """
+        # leave only last 2 dialogue turns to
+        # - capture previous meaningful action before action_listen
+        # - ignore previous intent
+        if len(states) == 1 or not states[-2].get(PREVIOUS_ACTION):
+            return [states[-1]]
+        else:
+            return [{PREVIOUS_ACTION: states[-2][PREVIOUS_ACTION]}, states[-1]]
+
+    @staticmethod
+    def _remove_rule_snippet_predictions(lookup: Dict[Text, Text]) -> Dict[Text, Text]:
+        # Delete rules if it would predict the RULE_SNIPPET_ACTION_NAME action
+        return {
+            feature_key: action
+            for feature_key, action in lookup.items()
+            if action != RULE_SNIPPET_ACTION_NAME
+        }
+
+    def _create_loop_unhappy_lookup_from_states(
+        self,
+        trackers_as_states: List[List[State]],
+        trackers_as_actions: List[List[Text]],
+    ) -> Dict[Text, Text]:
+        """Creates lookup dictionary from the tracker represented as states.
+
+        Args:
+            trackers_as_states: representation of the trackers as a list of states
+            trackers_as_actions: representation of the trackers as a list of actions
+
+        Returns:
+            lookup dictionary
+        """
+        lookup = {}
+        for states, actions in zip(trackers_as_states, trackers_as_actions):
+            action = actions[0]
+            active_loop = get_active_loop_name(states[-1])
+            # even if there are two identical feature keys
+            # their loop will be the same
+            if not active_loop:
+                continue
+
+            states = self._states_for_unhappy_loop_predictions(states)
+            feature_key = self._create_feature_key(states)
+            if not feature_key:
+                continue
+
+            # Since rule snippets and stories inside the loop contain
+            # only unhappy paths, notify the loop that
+            # it was predicted after an answer to a different question and
+            # therefore it should not validate user input
+            if (
+                # loop is predicted after action_listen in unhappy path,
+                # therefore no validation is needed
+                is_prev_action_listen_in_state(states[-1])
+                and action == active_loop
+            ):
+                lookup[feature_key] = LOOP_WAS_INTERRUPTED
+            elif (
+                # some action other than active_loop is predicted in unhappy path,
+                # therefore active_loop shouldn't be predicted by the rule
+                not is_prev_action_listen_in_state(states[-1])
+                and action != active_loop
+            ):
+                lookup[feature_key] = DO_NOT_PREDICT_LOOP_ACTION
+        return lookup
+
+    def _check_rule_restriction(
+        self, rule_trackers: List[TrackerWithCachedStates]
+    ) -> None:
+        rules_exceeding_max_user_turns = []
+        for tracker in rule_trackers:
+            number_of_user_uttered = sum(
+                isinstance(event, UserUttered) for event in tracker.events
+            )
+            if number_of_user_uttered > self.ALLOWED_NUMBER_OF_USER_INPUTS:
+                rules_exceeding_max_user_turns.append(tracker.sender_id)
+
+        if rules_exceeding_max_user_turns:
+            raise InvalidRule(
+                f"Found rules '{', '.join(rules_exceeding_max_user_turns)}' "
+                f"that contain more than {self.ALLOWED_NUMBER_OF_USER_INPUTS} "
+                f"user message. Rules are not meant to hardcode a state machine. "
+                f"Please use stories for these cases."
+            )
+
+    @staticmethod
+    def _expected_but_missing_slots(
+        fingerprint: ActionFingerprint, state: State
+    ) -> Set[Text]:
+        expected_slots = set(fingerprint.slots)
+        current_slots = set(state.get(SLOTS, {}).keys())
+        # report all slots that are expected but aren't set in current slots
+        return expected_slots.difference(current_slots)
+
+    @staticmethod
+    def _check_active_loops_fingerprint(
+        fingerprint: ActionFingerprint, state: State
+    ) -> Set[Optional[Text]]:
+        expected_active_loops = set(fingerprint.active_loop)
+        # we don't use tracker.active_loop_name
+        # because we need to keep should_not_be_set
+        current_active_loop = state.get(ACTIVE_LOOP, {}).get(LOOP_NAME)
+        if current_active_loop in expected_active_loops:
+            # one of expected active loops is set
+            return set()
+
+        return expected_active_loops
+
+    @staticmethod
+    def _error_messages_from_fingerprints(
+        action_name: Text,
+        missing_fingerprint_slots: Set[Text],
+        fingerprint_active_loops: Set[Text],
+        rule_name: Text,
+    ) -> List[Text]:
+        error_messages = []
+        if action_name and missing_fingerprint_slots:
+            error_messages.append(
+                f"- the action '{action_name}' in rule '{rule_name}' does not set some "
+                f"of the slots that it sets in other rules. Slots not set in rule "
+                f"'{rule_name}': '{', '.join(missing_fingerprint_slots)}'. Please "
+                f"update the rule with an appropriate slot or if it is the last action "
+                f"add 'wait_for_user_input: false' after this action."
+            )
+        if action_name and fingerprint_active_loops:
+            # substitute `SHOULD_NOT_BE_SET` with `null` so that users
+            # know what to put in their rules
+            fingerprint_active_loops = set(
+                "null" if active_loop == SHOULD_NOT_BE_SET else active_loop
+                for active_loop in fingerprint_active_loops
+            )
+            # add action_name to active loop so that users
+            # know what to put in their rules
+            fingerprint_active_loops.add(action_name)
+
+            error_messages.append(
+                f"- the form '{action_name}' in rule '{rule_name}' does not set "
+                f"the 'active_loop', that it sets in other rules: "
+                f"'{', '.join(fingerprint_active_loops)}'. Please update the rule with "
+                f"the appropriate 'active loop' property or if it is the last action "
+                f"add 'wait_for_user_input: false' after this action."
+            )
+        return error_messages
+
+    def _check_for_incomplete_rules(
+        self, rule_trackers: List[TrackerWithCachedStates], domain: Domain
+    ) -> None:
+        logger.debug("Started checking if some rules are incomplete.")
+        # we need to use only fingerprints from rules
+        rule_fingerprints = create_action_fingerprints(rule_trackers, domain)
+        if not rule_fingerprints:
+            return
+
+        error_messages: List[Text] = []
+        for tracker in rule_trackers:
+            states = tracker.past_states(domain)
+            # the last action is always action listen
+            action_names = [
+                state.get(PREVIOUS_ACTION, {}).get(ACTION_NAME) for state in states[1:]
+            ] + [ACTION_LISTEN_NAME]
+
+            for state, action_name in zip(states, action_names):
+                previous_action_name = state.get(PREVIOUS_ACTION, {}).get(ACTION_NAME)
+                fingerprint = rule_fingerprints.get(previous_action_name)
+                if (
+                    not previous_action_name
+                    or not fingerprint
+                    or action_name == RULE_SNIPPET_ACTION_NAME
+                    or previous_action_name == RULE_SNIPPET_ACTION_NAME
+                ):
+                    # do not check fingerprints for rule snippet action
+                    # and don't raise if fingerprints are not satisfied
+                    # for a previous action if current action is rule snippet action
+                    continue
+
+                missing_expected_slots = self._expected_but_missing_slots(
+                    fingerprint, state
+                )
+                expected_active_loops = self._check_active_loops_fingerprint(
+                    fingerprint, state
+                )
+                error_messages.extend(
+                    self._error_messages_from_fingerprints(
+                        previous_action_name,
+                        missing_expected_slots,
+                        expected_active_loops,
+                        tracker.sender_id,
+                    )
+                )
+
+        if error_messages:
+            error_text = "\n".join(error_messages)
+            raise InvalidRule(
+                f"\nIncomplete rules found🚨\n\n{error_text}\n"
+                f"Please note that if some slots or active loops should not be set "
+                f"during prediction you need to explicitly set them to 'null' in the "
+                f"rules."
+            )
+
+        logger.debug("Found no incompletions in rules.")
+
+    @staticmethod
+    def _get_slots_loops_from_states(
+        trackers_as_states: List[List[State]],
+    ) -> Tuple[Set[Text], Set[Text]]:
+        slots = set()
+        loops = set()
+        for states in trackers_as_states:
+            for state in states:
+                slots.update(set(state.get(SLOTS, {}).keys()))
+                # FIXME: ideally we have better annotation for State, TypedDict
+                # could work but support in mypy is very limited. Dataclass are
+                # another option
+                active_loop = cast(Text, state.get(ACTIVE_LOOP, {}).get(LOOP_NAME))
+                if active_loop:
+                    loops.add(active_loop)
+        return slots, loops
+
+    def _find_rule_only_slots_loops(
+        self,
+        rule_trackers_as_states: List[List[State]],
+        story_trackers_as_states: List[List[State]],
+    ) -> Tuple[List[Text], List[Text]]:
+        rule_slots, rule_loops = self._get_slots_loops_from_states(
+            rule_trackers_as_states
+        )
+        story_slots, story_loops = self._get_slots_loops_from_states(
+            story_trackers_as_states
+        )
+
+        # set is not json serializable, so convert to list
+        return (
+            list(rule_slots - story_slots - {SHOULD_NOT_BE_SET}),
+            list(rule_loops - story_loops - {SHOULD_NOT_BE_SET}),
+        )
+
+    def _predict_next_action(
+        self, tracker: TrackerWithCachedStates, domain: Domain
+    ) -> Tuple[Optional[Text], Optional[Text]]:
+        prediction, prediction_source = self._predict(tracker, domain)
+        probabilities = prediction.probabilities
+        # do not raise an error if RulePolicy didn't predict anything for stories;
+        # however for rules RulePolicy should always predict an action
+        predicted_action_name = None
+        if (
+            probabilities != self._default_predictions(domain)
+            or tracker.is_rule_tracker
+        ):
+            predicted_action_name = domain.action_names_or_texts[
+                np.argmax(probabilities)
+            ]
+
+        return predicted_action_name, prediction_source
+
+    def _predicted_action_name(
+        self, tracker: TrackerWithCachedStates, domain: Domain, gold_action_name: Text
+    ) -> Tuple[Optional[Text], Optional[Text]]:
+        predicted_action_name, prediction_source = self._predict_next_action(
+            tracker, domain
+        )
+        # if there is an active_loop,
+        # RulePolicy will always predict active_loop first,
+        # but inside loop unhappy path there might be another action
+        if (
+            tracker.active_loop_name
+            and predicted_action_name != gold_action_name
+            and predicted_action_name == tracker.active_loop_name
+        ):
+            rasa.core.test.emulate_loop_rejection(tracker)
+            predicted_action_name, prediction_source = self._predict_next_action(
+                tracker, domain
+            )
+
+        return predicted_action_name, prediction_source
+
+    def _collect_sources(
+        self,
+        tracker: TrackerWithCachedStates,
+        predicted_action_name: Optional[Text],
+        gold_action_name: Optional[Text],
+        prediction_source: Text,
+    ) -> None:
+        # we need to remember which action should be predicted by the rule
+        # in order to correctly output the names of the contradicting rules
+        rule_name = tracker.sender_id
+
+        if prediction_source is not None and (
+            prediction_source.startswith(DEFAULT_RULES)
+            or prediction_source.startswith(LOOP_RULES)
+        ):
+            # the real gold action contradict the one in the rules in this case
+            gold_action_name = predicted_action_name
+            rule_name = prediction_source
+
+        self._rules_sources[prediction_source].append((rule_name, gold_action_name))
+
+    @staticmethod
+    def _default_sources() -> Set[Text]:
+        return {
+            DEFAULT_RULES + default_intent
+            for default_intent in DEFAULT_ACTION_MAPPINGS.keys()
+        }
+
+    @staticmethod
+    def _handling_loop_sources(domain: Domain) -> Set[Text]:
+        loop_sources = set()
+        for loop_name in domain.form_names:
+            loop_sources.add(LOOP_RULES + loop_name)
+            loop_sources.add(
+                LOOP_RULES + loop_name + LOOP_RULES_SEPARATOR + ACTION_LISTEN_NAME
+            )
+        return loop_sources
+
+    def _should_delete(
+        self,
+        prediction_source: Text,
+        tracker: TrackerWithCachedStates,
+        predicted_action_name: Text,
+    ) -> bool:
+        """Checks whether this contradiction is due to action, intent pair.
+
+        Args:
+            prediction_source: the states that result in the prediction
+            tracker: the tracker that raises the contradiction
+            predicted_action_name: the action that was predicted
+
+        Returns:
+            true if the contradiction is a result of an action, intent pair in the rule.
+        """
+        if (
+            # only apply to contradicting story, not rule
+            tracker.is_rule_tracker
+            # only apply for prediction after unpredictable action
+            or prediction_source.count(PREVIOUS_ACTION) > 1
+            # only apply for prediction of action_listen
+            or predicted_action_name != ACTION_LISTEN_NAME
+        ):
+            return False
+        for source in self.lookup[RULES]:
+            # remove rule only if another action is predicted after action_listen
+            if (
+                source.startswith(prediction_source[:-2])
+                and not prediction_source == source
+            ):
+                return True
+        return False
+
+    def _check_prediction(
+        self,
+        tracker: TrackerWithCachedStates,
+        predicted_action_name: Optional[Text],
+        gold_action_name: Text,
+        prediction_source: Optional[Text],
+    ) -> List[Text]:
+        # FIXME: `predicted_action_name` and `prediction_source` are
+        # either None together or defined together. This could be improved
+        # by better typing in this class, but requires some refactoring
+        if (
+            not predicted_action_name
+            or not prediction_source
+            or predicted_action_name == gold_action_name
+        ):
+            return []
+
+        if self._should_delete(prediction_source, tracker, predicted_action_name):
+            self.lookup[RULES].pop(prediction_source)
+            return []
+
+        tracker_type = "rule" if tracker.is_rule_tracker else "story"
+        contradicting_rules = {
+            rule_name
+            for rule_name, action_name in self._rules_sources[prediction_source]
+            if action_name != gold_action_name
+        }
+
+        if not contradicting_rules:
+            return []
+
+        error_message = (
+            f"- the prediction of the action '{gold_action_name}' in {tracker_type} "
+            f"'{tracker.sender_id}' "
+            f"is contradicting with rule(s) '{', '.join(contradicting_rules)}'"
+        )
+        # outputting predicted action 'action_default_fallback' is confusing
+        if predicted_action_name != self._fallback_action_name:
+            error_message += f" which predicted action '{predicted_action_name}'"
+
+        return [error_message + "."]
+
+    def _run_prediction_on_trackers(
+        self,
+        trackers: List[TrackerWithCachedStates],
+        domain: Domain,
+        collect_sources: bool,
+    ) -> Tuple[List[Text], Set[Optional[Text]]]:
+        if collect_sources:
+            self._rules_sources = defaultdict(list)
+
+        error_messages = []
+        rules_used_in_stories = set()
+        pbar = tqdm(
+            trackers,
+            desc="Processed trackers",
+            disable=rasa.shared.utils.io.is_logging_disabled(),
+        )
+        for tracker in pbar:
+            running_tracker = tracker.init_copy()
+            running_tracker.sender_id = tracker.sender_id
+            # the first action is always unpredictable
+            next_action_is_unpredictable = True
+            for event in tracker.applied_events(True):
+                if not isinstance(event, ActionExecuted):
+                    running_tracker.update(event)
+                    continue
+
+                if event.action_name == RULE_SNIPPET_ACTION_NAME:
+                    # notify that the action after RULE_SNIPPET_ACTION_NAME is
+                    # unpredictable
+                    next_action_is_unpredictable = True
+                    running_tracker.update(event)
+                    continue
+
+                # do not run prediction on unpredictable actions
+                if next_action_is_unpredictable or event.unpredictable:
+                    next_action_is_unpredictable = False  # reset unpredictability
+                    running_tracker.update(event)
+                    continue
+
+                gold_action_name = event.action_name or event.action_text
+                predicted_action_name, prediction_source = self._predicted_action_name(
+                    running_tracker, domain, gold_action_name
+                )
+                if collect_sources:
+                    if prediction_source:
+                        self._collect_sources(
+                            running_tracker,
+                            predicted_action_name,
+                            gold_action_name,
+                            prediction_source,
+                        )
+                else:
+                    # to be able to remove only rules turns from the dialogue history
+                    # for ML policies,
+                    # we need to know which rules were used in ML trackers
+                    if (
+                        not tracker.is_rule_tracker
+                        and predicted_action_name == gold_action_name
+                    ):
+                        rules_used_in_stories.add(prediction_source)
+
+                    error_messages += self._check_prediction(
+                        running_tracker,
+                        predicted_action_name,
+                        gold_action_name,
+                        prediction_source,
+                    )
+
+                running_tracker.update(event)
+
+        return error_messages, rules_used_in_stories
+
+    def _collect_rule_sources(
+        self, rule_trackers: List[TrackerWithCachedStates], domain: Domain
+    ) -> None:
+        self._run_prediction_on_trackers(rule_trackers, domain, collect_sources=True)
+
+    def _find_contradicting_and_used_in_stories_rules(
+        self, trackers: List[TrackerWithCachedStates], domain: Domain
+    ) -> Tuple[List[Text], Set[Optional[Text]]]:
+        return self._run_prediction_on_trackers(trackers, domain, collect_sources=False)
+
+    def _analyze_rules(
+        self,
+        rule_trackers: List[TrackerWithCachedStates],
+        all_trackers: List[TrackerWithCachedStates],
+        domain: Domain,
+    ) -> List[Text]:
+        """Analyzes learned rules by running prediction on training trackers.
+
+        This method collects error messages for contradicting rules
+        and creates the lookup for rules that are not present in the stories.
+
+        Args:
+            rule_trackers: The list of the rule trackers.
+            all_trackers: The list of all trackers.
+            domain: The domain.
+
+        Returns:
+             Rules that are not present in the stories.
+        """
+        logger.debug("Started checking rules and stories for contradictions.")
+        # during training we run `predict_action_probabilities` to check for
+        # contradicting rules.
+        # We silent prediction debug to avoid too many logs during these checks.
+        logger_level = logger.level
+        logger.setLevel(logging.WARNING)
+
+        # we need to run prediction on rule trackers twice, because we need to collect
+        # the information about which rule snippets contributed to the learned rules
+        self._collect_rule_sources(rule_trackers, domain)
+        (
+            error_messages,
+            rules_used_in_stories,
+        ) = self._find_contradicting_and_used_in_stories_rules(all_trackers, domain)
+
+        logger.setLevel(logger_level)  # reset logger level
+        if error_messages:
+            error_text = "\n".join(error_messages)
+            raise InvalidRule(
+                f"\nContradicting rules or stories found 🚨\n\n{error_text}\n"
+                f"Please update your stories and rules so that they don't contradict "
+                f"each other."
+            )
+
+        logger.debug("Found no contradicting rules.")
+        all_rules = (
+            set(self._rules_sources.keys())
+            | self._default_sources()
+            | self._handling_loop_sources(domain)
+        )
+        # set is not json serializable, so convert to list
+        return list(all_rules - rules_used_in_stories)
+
+    def _create_lookup_from_trackers(
+        self,
+        rule_trackers: List[TrackerWithCachedStates],
+        story_trackers: List[TrackerWithCachedStates],
+        domain: Domain,
+    ) -> None:
+        (
+            rule_trackers_as_states,
+            rule_trackers_as_actions,
+        ) = self.featurizer.training_states_and_labels(
+            rule_trackers, domain, omit_unset_slots=True
+        )
+
+        rules_lookup = self._create_lookup_from_states(
+            rule_trackers_as_states, rule_trackers_as_actions
+        )
+        self.lookup[RULES] = self._remove_rule_snippet_predictions(rules_lookup)
+
+        (
+            story_trackers_as_states,
+            story_trackers_as_actions,
+        ) = self.featurizer.training_states_and_labels(story_trackers, domain)
+
+        if self._check_for_contradictions:
+            (
+                self.lookup[RULE_ONLY_SLOTS],
+                self.lookup[RULE_ONLY_LOOPS],
+            ) = self._find_rule_only_slots_loops(
+                rule_trackers_as_states, story_trackers_as_states
+            )
+
+        # use all trackers to find negative rules in unhappy paths
+        trackers_as_states = rule_trackers_as_states + story_trackers_as_states
+        trackers_as_actions = rule_trackers_as_actions + story_trackers_as_actions
+
+        # negative rules are not anti-rules, they are auxiliary to actual rules
+        self.lookup[
+            RULES_FOR_LOOP_UNHAPPY_PATH
+        ] = self._create_loop_unhappy_lookup_from_states(
+            trackers_as_states, trackers_as_actions
+        )
+
+    def train(
+        self,
+        training_trackers: List[TrackerWithCachedStates],
+        domain: Domain,
+        **kwargs: Any,
+    ) -> Resource:
+        """Trains the policy on given training trackers.
+
+        Args:
+            training_trackers: The list of the trackers.
+            domain: The domain.
+            **kwargs: Additional arguments.
+
+        Returns:
+            The resource which can be used to load the trained policy.
+        """
+        self.raise_if_incompatible_with_domain(self.config, domain)
+
+        # only consider original trackers (no augmented ones)
+        training_trackers = [
+            t for t in training_trackers if not getattr(t, "is_augmented", False)
+        ]
+        # trackers from rule-based training data
+        rule_trackers = [t for t in training_trackers if t.is_rule_tracker]
+        if self.config["restrict_rules"]:
+            self._check_rule_restriction(rule_trackers)
+        if self._check_for_contradictions:
+            self._check_for_incomplete_rules(rule_trackers, domain)
+
+        # trackers from ML-based training data
+        story_trackers = [t for t in training_trackers if not t.is_rule_tracker]
+
+        self._create_lookup_from_trackers(rule_trackers, story_trackers, domain)
+
+        # make this configurable because checking might take a lot of time
+        if self._check_for_contradictions:
+            # using trackers here might not be the most efficient way, however
+            # it allows us to directly test `predict_action_probabilities` method
+            self.lookup[RULES_NOT_IN_STORIES] = self._analyze_rules(
+                rule_trackers, training_trackers, domain
+            )
+
+        logger.debug(f"Memorized '{len(self.lookup[RULES])}' unique rules.")
+
+        self.persist()
+
+        return self._resource
+
+    @staticmethod
+    def _does_rule_match_state(rule_state: State, conversation_state: State) -> bool:
+        for state_type, rule_sub_state in rule_state.items():
+            conversation_sub_state = conversation_state.get(state_type, {})
+            for key, value_from_rules in rule_sub_state.items():
+                if isinstance(value_from_rules, list):
+                    # json dumps and loads tuples as lists,
+                    # so we need to convert them back
+                    value_from_rules = tuple(value_from_rules)
+                value_from_conversation = conversation_sub_state.get(key)
+                if (
+                    # value should be set, therefore
+                    # check whether it is the same as in the state
+                    value_from_rules
+                    and value_from_rules != SHOULD_NOT_BE_SET
+                    and value_from_conversation != value_from_rules
+                ) or (
+                    # value shouldn't be set, therefore
+                    # it should be None or non existent in the state
+                    value_from_rules == SHOULD_NOT_BE_SET
+                    and value_from_conversation
+                    # during training `SHOULD_NOT_BE_SET` is provided. Hence, we also
+                    # have to check for the value of the slot state
+                    and value_from_conversation != SHOULD_NOT_BE_SET
+                ):
+                    return False
+
+        return True
+
+    @staticmethod
+    # This function is called a lot (e.g. for checking contradictions) so we cache
+    # its results.
+    @functools.lru_cache(maxsize=1000)
+    def _rule_key_to_state(rule_key: Text) -> List[State]:
+        return json.loads(rule_key)
+
+    def _is_rule_applicable(
+        self, rule_key: Text, turn_index: int, conversation_state: State
+    ) -> bool:
+        """Checks if rule is satisfied with current state at turn.
+
+        Args:
+            rule_key: the textual representation of learned rule
+            turn_index: index of a current dialogue turn
+            conversation_state: the state that corresponds to turn_index
+
+        Returns:
+            a boolean that says whether the rule is applicable to current state
+        """
+        # turn_index goes back in time
+        reversed_rule_states = list(reversed(self._rule_key_to_state(rule_key)))
+
+        # the rule must be applicable because we got (without any applicability issues)
+        # further in the conversation history than the rule's length
+        if turn_index >= len(reversed_rule_states):
+            return True
+
+        # a state has previous action if and only if it is not a conversation start
+        # state
+        current_previous_action = conversation_state.get(PREVIOUS_ACTION)
+        rule_previous_action = reversed_rule_states[turn_index].get(PREVIOUS_ACTION)
+
+        # current conversation state and rule state are conversation starters.
+        # any slots with initial_value set will necessarily be in both states and don't
+        # need to be checked.
+        if not rule_previous_action and not current_previous_action:
+            return True
+
+        # current rule state is a conversation starter (due to conversation_start: true)
+        # but current conversation state is not.
+        # or
+        # current conversation state is a starter
+        # but current rule state is not.
+        if not rule_previous_action or not current_previous_action:
+            return False
+
+        # check: current rule state features are present in current conversation state
+        return self._does_rule_match_state(
+            reversed_rule_states[turn_index], conversation_state
+        )
+
+    def _get_possible_keys(
+        self, lookup: Dict[Text, Text], states: List[State]
+    ) -> Set[Text]:
+        possible_keys = set(lookup.keys())
+        for i, state in enumerate(reversed(states)):
+            # find rule keys that correspond to current state
+            possible_keys = set(
+                filter(
+                    lambda _key: self._is_rule_applicable(_key, i, state), possible_keys
+                )
+            )
+        return possible_keys
+
+    @staticmethod
+    def _find_action_from_default_actions(
+        tracker: DialogueStateTracker,
+    ) -> Tuple[Optional[Text], Optional[Text]]:
+        if (
+            not tracker.latest_action_name == ACTION_LISTEN_NAME
+            or not tracker.latest_message
+        ):
+            return None, None
+
+        intent_name = tracker.latest_message.intent.get(INTENT_NAME_KEY)
+        if intent_name is None:
+            return None, None
+
+        default_action_name = DEFAULT_ACTION_MAPPINGS.get(intent_name)
+        if default_action_name is None:
+            return None, None
+
+        logger.debug(f"Predicted default action '{default_action_name}'.")
+        return (
+            default_action_name,
+            # create prediction source that corresponds to one of
+            # default prediction sources in `_default_sources()`
+            DEFAULT_RULES + intent_name,
+        )
+
+    @staticmethod
+    def _find_action_from_loop_happy_path(
+        tracker: DialogueStateTracker,
+    ) -> Tuple[Optional[Text], Optional[Text]]:
+
+        active_loop_name = tracker.active_loop_name
+        if active_loop_name is None:
+            return None, None
+
+        active_loop_rejected = tracker.is_active_loop_rejected
+        should_predict_loop = (
+            not active_loop_rejected
+            and tracker.latest_action
+            and tracker.latest_action.get(ACTION_NAME) != active_loop_name
+        )
+        should_predict_listen = (
+            not active_loop_rejected and tracker.latest_action_name == active_loop_name
+        )
+
+        if should_predict_loop:
+            logger.debug(f"Predicted loop '{active_loop_name}'.")
+            return active_loop_name, LOOP_RULES + active_loop_name
+
+        # predict `action_listen` if loop action was run successfully
+        if should_predict_listen:
+            logger.debug(
+                f"Predicted '{ACTION_LISTEN_NAME}' after loop '{active_loop_name}'."
+            )
+            return (
+                ACTION_LISTEN_NAME,
+                (
+                    f"{LOOP_RULES}{active_loop_name}"
+                    f"{LOOP_RULES_SEPARATOR}{ACTION_LISTEN_NAME}"
+                ),
+            )
+
+        return None, None
+
+    def _find_action_from_rules(
+        self,
+        tracker: DialogueStateTracker,
+        domain: Domain,
+        use_text_for_last_user_input: bool,
+    ) -> Tuple[Optional[Text], Optional[Text], bool]:
+        """Predicts the next action based on the memoized rules.
+
+        Args:
+            tracker: The current conversation tracker.
+            domain: The domain of the current model.
+            use_text_for_last_user_input: `True` if text of last user message
+                should be used for the prediction. `False` if intent should be used.
+
+        Returns:
+            A tuple of the predicted action name or text (or `None` if no matching rule
+            was found), a description of the matching rule, and `True` if a loop action
+            was predicted after the loop has been in an unhappy path before.
+        """
+        if (
+            use_text_for_last_user_input
+            and not tracker.latest_action_name == ACTION_LISTEN_NAME
+        ):
+            # make text prediction only directly after user utterance
+            # because we've otherwise already decided whether to use
+            # the text or the intent
+            return None, None, False
+
+        states = self._prediction_states(
+            tracker,
+            domain,
+            use_text_for_last_user_input,
+            rule_only_data=self._get_rule_only_data(),
+        )
+
+        current_states = self.format_tracker_states(states)
+        structlogger.debug(
+            "rule_policy.actions.find", current_states=copy.deepcopy(current_states)
+        )
+
+        # Tracks if we are returning after an unhappy loop path. If this becomes `True`
+        # the policy returns an event which notifies the loop action that it
+        # is returning after an unhappy path. For example, the `FormAction` uses this
+        # to skip the validation of slots for its first execution after an unhappy path.
+        returning_from_unhappy_path = False
+
+        rule_keys = self._get_possible_keys(self.lookup[RULES], states)
+        predicted_action_name = None
+        best_rule_key = ""
+        if rule_keys:
+            # if there are several rules,
+            # it should mean that some rule is a subset of another rule
+            # therefore we pick a rule of maximum length
+            best_rule_key = max(rule_keys, key=len)
+            predicted_action_name = self.lookup[RULES].get(best_rule_key)
+
+        active_loop_name = tracker.active_loop_name
+        if active_loop_name:
+            # find rules for unhappy path of the loop
+            loop_unhappy_keys = self._get_possible_keys(
+                self.lookup[RULES_FOR_LOOP_UNHAPPY_PATH], states
+            )
+            # there could be several unhappy path conditions
+            unhappy_path_conditions = [
+                self.lookup[RULES_FOR_LOOP_UNHAPPY_PATH].get(key)
+                for key in loop_unhappy_keys
+            ]
+
+            # Check if a rule that predicted action_listen
+            # was applied inside the loop.
+            # Rules might not explicitly switch back to the loop.
+            # Hence, we have to take care of that.
+            predicted_listen_from_general_rule = (
+                predicted_action_name == ACTION_LISTEN_NAME
+                and not get_active_loop_name(self._rule_key_to_state(best_rule_key)[-1])
+            )
+            if predicted_listen_from_general_rule:
+                if DO_NOT_PREDICT_LOOP_ACTION not in unhappy_path_conditions:
+                    # negative rules don't contain a key that corresponds to
+                    # the fact that active_loop shouldn't be predicted
+                    logger.debug(
+                        f"Predicted loop '{active_loop_name}' by overwriting "
+                        f"'{ACTION_LISTEN_NAME}' predicted by general rule."
+                    )
+                    return (
+                        active_loop_name,
+                        best_rule_key,
+                        returning_from_unhappy_path,
+                    )
+
+                # do not predict anything
+                predicted_action_name = None
+
+            if LOOP_WAS_INTERRUPTED in unhappy_path_conditions:
+                logger.debug(
+                    "Returning from unhappy path. Loop will be notified that "
+                    "it was interrupted."
+                )
+                returning_from_unhappy_path = True
+
+        if predicted_action_name is not None:
+            logger.debug(
+                f"There is a rule for the next action '{predicted_action_name}'."
+            )
+        else:
+            logger.debug("There is no applicable rule.")
+
+        # if we didn't predict anything from the rules, then the feature key created
+        # from states can be used as an indicator that this state will lead to fallback
+        return (
+            predicted_action_name,
+            best_rule_key or self._create_feature_key(states),
+            returning_from_unhappy_path,
+        )
+
+    async def predict_action_probabilities(
+        self,
+        tracker: DialogueStateTracker,
+        domain: Domain,
+        rule_only_data: Optional[Dict[Text, Any]] = None,
+        **kwargs: Any,
+    ) -> PolicyPrediction:
+        """Predicts the next action (see parent class for more information)."""
+        if self.should_abstain_in_coexistence(tracker, False):
+            # don't use self._default_predictions as this might have a different
+            # probability for the fallback action, and we want to have all probabilities
+            # set to 0.0
+            return self._prediction(super()._default_predictions(domain))
+
+        prediction, _ = self._predict(tracker, domain)
+        return prediction
+
+    def _predict(
+        self, tracker: DialogueStateTracker, domain: Domain
+    ) -> Tuple[PolicyPrediction, Optional[Text]]:
+        (
+            rules_action_name_from_text,
+            prediction_source_from_text,
+            returning_from_unhappy_path_from_text,
+        ) = self._find_action_from_rules(
+            tracker, domain, use_text_for_last_user_input=True
+        )
+
+        # Rasa Open Source default actions overrule anything. If users want to achieve
+        # the same, they need to write a rule or make sure that their loop rejects
+        # accordingly.
+        (
+            default_action_name,
+            default_prediction_source,
+        ) = self._find_action_from_default_actions(tracker)
+
+        # text has priority over intents including default,
+        # however loop happy path has priority over rules prediction
+        if default_action_name and not rules_action_name_from_text:
+            return (
+                self._rule_prediction(
+                    self._prediction_result(default_action_name, tracker, domain),
+                    default_prediction_source,
+                ),
+                default_prediction_source,
+            )
+
+        # A loop has priority over any other rule except defaults.
+        # The rules or any other prediction will be applied only if a loop was rejected.
+        # If we are in a loop, and the loop didn't run previously or rejected, we can
+        # simply force predict the loop.
+        (
+            loop_happy_path_action_name,
+            loop_happy_path_prediction_source,
+        ) = self._find_action_from_loop_happy_path(tracker)
+        if loop_happy_path_action_name:
+            # this prediction doesn't use user input
+            # and happy user input anyhow should be ignored during featurization
+            return (
+                self._rule_prediction(
+                    self._prediction_result(
+                        loop_happy_path_action_name, tracker, domain
+                    ),
+                    loop_happy_path_prediction_source,
+                    is_no_user_prediction=True,
+                ),
+                loop_happy_path_prediction_source,
+            )
+
+        # predict rules from text first
+        if rules_action_name_from_text:
+            return (
+                self._rule_prediction(
+                    self._prediction_result(
+                        rules_action_name_from_text, tracker, domain
+                    ),
+                    prediction_source_from_text,
+                    returning_from_unhappy_path=returning_from_unhappy_path_from_text,
+                    is_end_to_end_prediction=True,
+                ),
+                prediction_source_from_text,
+            )
+
+        (
+            rules_action_name_from_intent,
+            # we want to remember the source even if rules didn't predict any action
+            prediction_source_from_intent,
+            returning_from_unhappy_path_from_intent,
+        ) = self._find_action_from_rules(
+            tracker, domain, use_text_for_last_user_input=False
+        )
+        if rules_action_name_from_intent:
+            probabilities = self._prediction_result(
+                rules_action_name_from_intent, tracker, domain
+            )
+        else:
+            probabilities = self._default_predictions(domain)
+
+        return (
+            self._rule_prediction(
+                probabilities,
+                prediction_source_from_intent,
+                returning_from_unhappy_path=(
+                    # returning_from_unhappy_path is a negative condition,
+                    # so `or` should be applied
+                    returning_from_unhappy_path_from_text
+                    or returning_from_unhappy_path_from_intent
+                ),
+                is_end_to_end_prediction=False,
+            ),
+            prediction_source_from_intent,
+        )
+
+    def _rule_prediction(
+        self,
+        probabilities: List[float],
+        prediction_source: Text,
+        returning_from_unhappy_path: bool = False,
+        is_end_to_end_prediction: bool = False,
+        is_no_user_prediction: bool = False,
+    ) -> PolicyPrediction:
+        return PolicyPrediction(
+            probabilities,
+            self.__class__.__name__,
+            self.priority,
+            events=[LoopInterrupted(True)] if returning_from_unhappy_path else [],
+            is_end_to_end_prediction=is_end_to_end_prediction,
+            is_no_user_prediction=is_no_user_prediction,
+            hide_rule_turn=(
+                True
+                if prediction_source in self.lookup.get(RULES_NOT_IN_STORIES, [])
+                else False
+            ),
+        )
+
+    def _default_predictions(self, domain: Domain) -> List[float]:
+        result = super()._default_predictions(domain)
+
+        if self._enable_fallback_prediction:
+            result[domain.index_for_action(self._fallback_action_name)] = self.config[
+                "core_fallback_threshold"
+            ]
+
+        return result
+
+    def persist(self) -> None:
+        """Persists trained `RulePolicy`."""
+        super().persist()
+        with self._model_storage.write_to(self._resource) as directory:
+            rule_only_data = self._get_rule_only_data()
+            rasa.shared.utils.io.dump_obj_as_json_to_file(
+                directory / "rule_only_data.json", rule_only_data
+            )
+
+    def _metadata(self) -> Dict[Text, Any]:
+        return {"lookup": self.lookup}
+
+    @classmethod
+    def _metadata_filename(cls) -> Text:
+        return "rule_policy.json"
+
+    def _get_rule_only_data(self) -> Dict[Text, Any]:
+        """Gets the slots and loops that are used only in rule data.
+
+        Returns:
+            Slots and loops that are used only in rule data.
+        """
+        return {
+            key: self.lookup.get(key, []) for key in [RULE_ONLY_SLOTS, RULE_ONLY_LOOPS]
+        }