rasa-pro 3.12.0.dev1__py3-none-any.whl

rasa/dialogue_understanding/generator/multi_step/multi_step_llm_command_generator.py

@@ -0,0 +1,920 @@
+import importlib.resources
+import re
+from typing import Dict, Any, List, Optional, Tuple, Union, Text
+
+import structlog
+from jinja2 import Template
+
+import rasa.shared.utils.io
+from rasa.dialogue_understanding.commands import (
+    Command,
+    ErrorCommand,
+    SetSlotCommand,
+    CancelFlowCommand,
+    StartFlowCommand,
+    HumanHandoffCommand,
+    ChitChatAnswerCommand,
+    SkipQuestionCommand,
+    KnowledgeAnswerCommand,
+    ClarifyCommand,
+    CannotHandleCommand,
+)
+from rasa.dialogue_understanding.commands.change_flow_command import ChangeFlowCommand
+from rasa.dialogue_understanding.generator.constants import (
+    LLM_CONFIG_KEY,
+    USER_INPUT_CONFIG_KEY,
+    FLOW_RETRIEVAL_KEY,
+    DEFAULT_LLM_CONFIG,
+)
+from rasa.dialogue_understanding.generator.flow_retrieval import FlowRetrieval
+from rasa.dialogue_understanding.generator.llm_based_command_generator import (
+    LLMBasedCommandGenerator,
+)
+from rasa.dialogue_understanding.stack.frames import UserFlowStackFrame
+from rasa.dialogue_understanding.stack.utils import (
+    top_flow_frame,
+    top_user_flow_frame,
+    user_flows_on_the_stack,
+)
+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 (
+    RASA_PATTERN_CANNOT_HANDLE_NOT_SUPPORTED,
+    EMBEDDINGS_CONFIG_KEY,
+)
+from rasa.shared.constants import ROUTE_TO_CALM_SLOT
+from rasa.shared.core.flows import FlowStep, Flow, FlowsList
+from rasa.shared.core.flows.steps.collect import CollectInformationFlowStep
+from rasa.shared.core.trackers import DialogueStateTracker
+from rasa.shared.exceptions import ProviderClientAPIException
+from rasa.shared.nlu.constants import TEXT
+from rasa.shared.nlu.training_data.message import Message
+from rasa.shared.utils.io import deep_container_fingerprint
+from rasa.shared.utils.llm import (
+    get_prompt_template,
+    tracker_as_readable_transcript,
+    sanitize_message_for_prompt,
+    allowed_values_for_slot,
+    resolve_model_client_config,
+)
+
+# multistep template keys
+HANDLE_FLOWS_KEY = "handle_flows"
+FILL_SLOTS_KEY = "fill_slots"
+
+# multistep template file names
+HANDLE_FLOWS_PROMPT_FILE_NAME = "handle_flows_prompt.jinja2"
+FILL_SLOTS_PROMPT_FILE_NAME = "fill_slots_prompt.jinja2"
+
+# multistep templates
+DEFAULT_HANDLE_FLOWS_TEMPLATE = importlib.resources.read_text(
+    "rasa.dialogue_understanding.generator.multi_step", "handle_flows_prompt.jinja2"
+).strip()
+DEFAULT_FILL_SLOTS_TEMPLATE = importlib.resources.read_text(
+    "rasa.dialogue_understanding.generator.multi_step", "fill_slots_prompt.jinja2"
+).strip()
+MULTI_STEP_LLM_COMMAND_GENERATOR_CONFIG_FILE = "config.json"
+
+# dictionary of template names and associated file names and default values
+PROMPT_TEMPLATES = {
+    HANDLE_FLOWS_KEY: (
+        HANDLE_FLOWS_PROMPT_FILE_NAME,
+        DEFAULT_HANDLE_FLOWS_TEMPLATE,
+    ),
+    FILL_SLOTS_KEY: (
+        FILL_SLOTS_PROMPT_FILE_NAME,
+        DEFAULT_FILL_SLOTS_TEMPLATE,
+    ),
+}
+
+FILE_PATH_KEY = "file_path"
+
+structlogger = structlog.get_logger()
+
+
+@DefaultV1Recipe.register(
+    [
+        DefaultV1Recipe.ComponentType.COMMAND_GENERATOR,
+    ],
+    is_trainable=True,
+)
+class MultiStepLLMCommandGenerator(LLMBasedCommandGenerator):
+    """An multi step command generator using LLM."""
+
+    def __init__(
+        self,
+        config: Dict[Text, Any],
+        model_storage: ModelStorage,
+        resource: Resource,
+        prompt_templates: Optional[Dict[Text, Optional[Text]]] = None,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(
+            config, model_storage, resource, prompt_templates=prompt_templates, **kwargs
+        )
+
+        self._prompts: Dict[Text, Optional[Text]] = {
+            HANDLE_FLOWS_KEY: None,
+            FILL_SLOTS_KEY: None,
+        }
+        self._init_prompt_templates(prompt_templates)
+
+        self.trace_prompt_tokens = self.config.get("trace_prompt_tokens", False)
+
+    ### Implementations of LLMBasedCommandGenerator parent
+    @staticmethod
+    def get_default_config() -> Dict[str, Any]:
+        """The component's default config (see parent class for full docstring)."""
+        return {
+            "prompt_templates": {},
+            USER_INPUT_CONFIG_KEY: None,
+            LLM_CONFIG_KEY: None,
+            FLOW_RETRIEVAL_KEY: FlowRetrieval.get_default_config(),
+        }
+
+    @classmethod
+    def load(
+        cls: Any,
+        config: Dict[str, Any],
+        model_storage: ModelStorage,
+        resource: Resource,
+        execution_context: ExecutionContext,
+        **kwargs: Any,
+    ) -> "MultiStepLLMCommandGenerator":
+        """Loads trained component (see parent class for full docstring)."""
+        # Perform health check of the LLM client config
+        llm_config = resolve_model_client_config(config.get(LLM_CONFIG_KEY, {}))
+        cls.perform_llm_health_check(
+            llm_config,
+            DEFAULT_LLM_CONFIG,
+            "multi_step_llm_command_generator.load",
+            MultiStepLLMCommandGenerator.__name__,
+        )
+
+        prompts = cls._load_prompt_templates(model_storage, resource)
+
+        # init base command generator
+        command_generator = cls(config, model_storage, resource, prompts)
+        # load flow retrieval if enabled
+        if command_generator.enabled_flow_retrieval:
+            command_generator.flow_retrieval = cls.load_flow_retrival(
+                command_generator.config, model_storage, resource
+            )
+
+        return command_generator
+
+    def persist(self) -> None:
+        """Persist this component to disk for future loading."""
+        self._persist_prompt_templates()
+        self._persist_config()
+        if self.flow_retrieval is not None:
+            self.flow_retrieval.persist()
+
+    async def predict_commands(
+        self,
+        message: Message,
+        flows: FlowsList,
+        tracker: Optional[DialogueStateTracker] = None,
+        **kwargs: Any,
+    ) -> List[Command]:
+        """Predict commands using the LLM.
+
+        Args:
+            message: The message from the user.
+            flows: The flows available to the user.
+            tracker: The tracker containing the current state of the conversation.
+            **kwargs: Keyword arguments for forward compatibility.
+
+        Returns:
+            The commands generated by the llm.
+        """
+        if tracker is None or flows.is_empty():
+            # cannot do anything if there are no flows or no tracker
+            return []
+
+        try:
+            commands = await self._predict_commands_with_multi_step(
+                message, flows, tracker
+            )
+            commands = self._clean_up_commands(commands)
+            self._add_commands_to_message_parse_data(
+                message, MultiStepLLMCommandGenerator.__name__, commands
+            )
+        except ProviderClientAPIException:
+            # if any step resulted in API exception, the command prediction cannot
+            # be completed, "predict" the ErrorCommand
+            commands = [ErrorCommand()]
+
+        if not commands:
+            # if for any reason the final list of commands is empty,
+            # "predict" CannotHandle
+            commands = [CannotHandleCommand()]
+
+        if tracker.has_coexistence_routing_slot:
+            # if coexistence feature is used, set the routing slot
+            commands += [SetSlotCommand(ROUTE_TO_CALM_SLOT, True)]
+
+        structlogger.debug(
+            "multi_step_llm_command_generator.predict_commands.finished",
+            commands=commands,
+        )
+
+        return commands
+
+    @classmethod
+    def parse_commands(
+        cls,
+        actions: Optional[str],
+        tracker: DialogueStateTracker,
+        flows: FlowsList,
+        is_handle_flows_prompt: bool = False,
+    ) -> List[Command]:
+        """Parse the actions returned by the llm into intent and entities.
+
+        Args:
+            actions: The actions returned by the llm.
+            tracker: The tracker containing the current state of the conversation.
+            flows: The list of flows.
+            is_handle_flows_prompt: bool
+
+        Returns:
+            The parsed commands.
+        """
+        if not actions:
+            return []
+
+        commands: List[Command] = []
+
+        slot_set_re = re.compile(
+            r"""SetSlot\(['"]?([a-zA-Z_][a-zA-Z0-9_-]*)['"]?, ?['"]?(.*)['"]?\)"""
+        )
+        start_flow_re = re.compile(r"StartFlow\(['\"]?([a-zA-Z0-9_-]+)['\"]?\)")
+        change_flow_re = re.compile(r"ChangeFlow\(\)")
+        cancel_flow_re = re.compile(r"CancelFlow\(\)")
+        chitchat_re = re.compile(r"ChitChat\(\)")
+        skip_question_re = re.compile(r"SkipQuestion\(\)")
+        knowledge_re = re.compile(r"SearchAndReply\(\)")
+        humand_handoff_re = re.compile(r"HumanHandoff\(\)")
+        clarify_re = re.compile(r"Clarify\(([\"\'a-zA-Z0-9_, ]+)\)")
+        cannot_handle_re = re.compile(r"CannotHandle\(\)")
+
+        for action in actions.strip().splitlines():
+            if is_handle_flows_prompt:
+                if (
+                    len(commands) >= 2
+                    or len(commands) == 1
+                    and isinstance(commands[0], ClarifyCommand)
+                ):
+                    break
+
+            if cannot_handle_re.search(action):
+                commands.append(
+                    CannotHandleCommand(RASA_PATTERN_CANNOT_HANDLE_NOT_SUPPORTED)
+                )
+            if match := slot_set_re.search(action):
+                slot_name = cls.clean_extracted_value(match.group(1).strip())
+                slot_value = cls.clean_extracted_value(match.group(2))
+                # error case where the llm tries to start a flow using a slot set
+                if slot_name == "flow_name":
+                    commands.extend(cls.start_flow_by_name(slot_value, flows))
+                else:
+                    typed_slot_value = cls.get_nullable_slot_value(slot_value)
+                    commands.append(
+                        SetSlotCommand(name=slot_name, value=typed_slot_value)
+                    )
+            elif match := start_flow_re.search(action):
+                flow_name = match.group(1).strip()
+                commands.extend(cls.start_flow_by_name(flow_name, flows))
+            elif cancel_flow_re.search(action):
+                commands.append(CancelFlowCommand())
+            elif chitchat_re.search(action):
+                commands.append(ChitChatAnswerCommand())
+            elif skip_question_re.search(action):
+                commands.append(SkipQuestionCommand())
+            elif knowledge_re.search(action):
+                commands.append(KnowledgeAnswerCommand())
+            elif humand_handoff_re.search(action):
+                commands.append(HumanHandoffCommand())
+            elif match := clarify_re.search(action):
+                options = sorted([opt.strip() for opt in match.group(1).split(",")])
+                # Remove surrounding quotes if present
+                cleaned_options = []
+                for flow in options:
+                    if (flow.startswith('"') and flow.endswith('"')) or (
+                        flow.startswith("'") and flow.endswith("'")
+                    ):
+                        cleaned_options.append(flow[1:-1])
+                    else:
+                        cleaned_options.append(flow)
+                # check if flow is valid
+                valid_options = [
+                    flow
+                    for flow in cleaned_options
+                    if flow in flows.user_flow_ids
+                    and flow not in user_flows_on_the_stack(tracker.stack)
+                ]
+                if len(valid_options) == 1:
+                    commands.extend(cls.start_flow_by_name(valid_options[0], flows))
+                elif 1 < len(valid_options) <= 5:
+                    commands.append(ClarifyCommand(valid_options))
+            elif change_flow_re.search(action):
+                commands.append(ChangeFlowCommand())
+
+        if not commands:
+            structlogger.debug(
+                "multi_step_llm_command_generator.parse_commands",
+                message="No commands were parsed from the LLM actions.",
+                actions=actions,
+            )
+
+        return commands
+
+    ### Helper methods
+    @property
+    def handle_flows_prompt(self) -> Optional[Text]:
+        return self._prompts[HANDLE_FLOWS_KEY]
+
+    @property
+    def fill_slots_prompt(self) -> Optional[Text]:
+        return self._prompts[FILL_SLOTS_KEY]
+
+    def _init_prompt_templates(self, prompt_templates: Dict[Text, Any]) -> None:
+        for key in self._prompts.keys():
+            _, default_template = PROMPT_TEMPLATES[key]
+            self._prompts[key] = self._resolve_prompt_template(
+                prompt_templates, self.config, key, default_template
+            )
+
+    @staticmethod
+    def _resolve_prompt_template(
+        prompt_templates: Optional[Dict[Text, Optional[Text]]],
+        config: Dict[Text, Any],
+        key: Text,
+        default_value: Text,
+    ) -> Text:
+        """Determines and retrieves a prompt template for a specific step in the
+        multistep command generator process using a given key. If the prompt
+        associated with the key is missing in both the `prompt_templates` and the
+        `config`, this method defaults to using a predefined prompt template. Each key
+        is uniquely associated with a distinct step of the command generation process.
+
+        Args:
+            prompt_templates: A dictionary of override templates.
+            config: The components config that may contain the file paths to the prompt
+            templates.
+            key: The key for the desired template.
+            default_value: The default template to use if no other is found.
+
+        Returns:
+            Prompt template.
+        """
+        if (
+            prompt_templates is not None
+            and key in prompt_templates
+            and prompt_templates[key] is not None
+        ):
+            return prompt_templates[key]  # type: ignore[return-value]
+        return get_prompt_template(
+            config.get("prompt_templates", {}).get(key, {}).get(FILE_PATH_KEY),
+            default_value,
+        )
+
+    @classmethod
+    def _load_prompt_templates(
+        cls, model_storage: ModelStorage, resource: Resource
+    ) -> Dict[Text, Text]:
+        """Loads persisted prompt templates from the model storage. If a prompt template
+        cannot be loaded, default value is used.
+        """
+        prompts = {}
+        for key, (file_name, default_value) in PROMPT_TEMPLATES.items():
+            prompt_template = cls.load_prompt_template_from_model_storage(
+                model_storage, resource, file_name
+            )
+            prompts[key] = prompt_template if prompt_template else default_value
+        return prompts
+
+    def _persist_prompt_templates(self) -> None:
+        """Persist the prompt templates to disk for future loading."""
+        with self._model_storage.write_to(self._resource) as path:
+            for key, template in self._prompts.items():
+                file_name, _ = PROMPT_TEMPLATES[key]
+                file_path = path / file_name
+                rasa.shared.utils.io.write_text_file(template, file_path)
+
+    def _persist_config(self) -> None:
+        """Persist config as a source of truth for resolved clients."""
+        with self._model_storage.write_to(self._resource) as path:
+            rasa.shared.utils.io.dump_obj_as_json_to_file(
+                path / MULTI_STEP_LLM_COMMAND_GENERATOR_CONFIG_FILE, self.config
+            )
+
+    async def _predict_commands_with_multi_step(
+        self,
+        message: Message,
+        flows: FlowsList,
+        tracker: DialogueStateTracker,
+    ) -> List[Command]:
+        """Predict commands using the LLM.
+
+        Args:
+            message: The message from the user.
+            flows: The flows available to the user.
+            tracker: The tracker containing the current state of the conversation.
+
+        Returns:
+            The commands generated by the llm.
+
+        Raises:
+            ProviderClientAPIException: If API calls raised an error.
+        """
+        # retrieve relevant flows
+        filtered_flows = await self.filter_flows(message, flows, tracker)
+
+        # 1st step: Handle active flow
+        if tracker.has_active_user_flow:
+            commands_from_active_flow = await self._predict_commands_for_active_flow(
+                message,
+                tracker,
+                available_flows=filtered_flows,
+                all_flows=flows,
+            )
+        else:
+            commands_from_active_flow = []
+
+        # 2nd step: Check if we need to switch to another flow
+        contains_change_flow_command = any(
+            isinstance(command, ChangeFlowCommand)
+            for command in commands_from_active_flow
+        )
+        should_change_flows = (
+            not commands_from_active_flow or contains_change_flow_command
+        )
+
+        if should_change_flows:
+            commands_for_handling_flows = (
+                await self._predict_commands_for_handling_flows(
+                    message,
+                    tracker,
+                    available_flows=filtered_flows,
+                    all_flows=flows,
+                )
+            )
+        else:
+            commands_for_handling_flows = []
+
+        if contains_change_flow_command:
+            commands_from_active_flow.pop(
+                commands_from_active_flow.index(ChangeFlowCommand())
+            )
+
+        # 3rd step: Fill slots for started flows
+        newly_started_flows = FlowsList(
+            [
+                flow
+                for command in commands_for_handling_flows
+                if (
+                    isinstance(command, StartFlowCommand)
+                    and (flow := filtered_flows.flow_by_id(command.flow)) is not None
+                )
+            ]
+        )
+
+        commands_for_newly_started_flows = (
+            await self._predict_commands_for_newly_started_flows(
+                message,
+                tracker,
+                newly_started_flows=newly_started_flows,
+                all_flows=flows,
+            )
+        )
+
+        # concatenate predicted commands
+        commands = list(
+            set(
+                commands_from_active_flow
+                + commands_for_handling_flows
+                + commands_for_newly_started_flows
+            )
+        )
+
+        return commands
+
+    async def _predict_commands_for_active_flow(
+        self,
+        message: Message,
+        tracker: DialogueStateTracker,
+        available_flows: FlowsList,
+        all_flows: FlowsList,
+    ) -> List[Command]:
+        """Predicts set slots commands for currently active flow.
+
+        Args:
+            message: The message from the user.
+            tracker: The tracker containing the current state of the conversation.
+            available_flows: Startable and active flows.
+            all_flows: All flows.
+
+        inputs = self._prepare_inputs(message, tracker, startable_flows, all_flows)
+
+        Returns:
+            Predicted commands for the active flow.
+        """
+        inputs = self._prepare_inputs(message, tracker, available_flows, all_flows)
+
+        if inputs["current_flow"] is None:
+            return []
+
+        prompt = Template(self.fill_slots_prompt).render(**inputs).strip()
+        structlogger.debug(
+            "multi_step_llm_command_generator"
+            ".predict_commands_for_active_flow"
+            ".prompt_rendered",
+            prompt=prompt,
+        )
+
+        actions = await self.invoke_llm(prompt)
+        structlogger.debug(
+            "multi_step_llm_command_generator"
+            ".predict_commands_for_active_flow"
+            ".actions_generated",
+            action_list=actions,
+        )
+
+        commands = self.parse_commands(actions, tracker, available_flows)
+
+        if commands:
+            self._add_prompt_to_message_parse_data(
+                message,
+                MultiStepLLMCommandGenerator.__name__,
+                "fill_slots_for_active_flow_prompt",
+                prompt,
+            )
+
+        return commands
+
+    async def _predict_commands_for_handling_flows(
+        self,
+        message: Message,
+        tracker: DialogueStateTracker,
+        available_flows: FlowsList,
+        all_flows: FlowsList,
+    ) -> List[Command]:
+        """Predicts commands for starting and canceling flows.
+
+        Args:
+            message: The message from the user.
+            tracker: The tracker containing the current state of the conversation.
+            available_flows: Startable and active flows.
+            all_flows: All flows.
+
+        inputs = self._prepare_inputs(message, tracker, startable_flows, all_flows, 2)
+
+        Returns:
+            Predicted commands for the starting/canceling flows.
+        """
+        inputs = self._prepare_inputs(message, tracker, available_flows, all_flows, 2)
+        prompt = Template(self.handle_flows_prompt).render(**inputs).strip()
+        structlogger.debug(
+            "multi_step_llm_command_generator"
+            ".predict_commands_for_handling_flows"
+            ".prompt_rendered",
+            prompt=prompt,
+        )
+
+        actions = await self.invoke_llm(prompt)
+        structlogger.debug(
+            "multi_step_llm_command_generator"
+            ".predict_commands_for_handling_flows"
+            ".actions_generated",
+            action_list=actions,
+        )
+
+        commands = self.parse_commands(actions, tracker, available_flows, True)
+        # filter out flows that are already started and active
+        commands = self._filter_redundant_start_flow_commands(tracker, commands)
+
+        if commands:
+            self._add_prompt_to_message_parse_data(
+                message,
+                MultiStepLLMCommandGenerator.__name__,
+                "handle_flows_prompt",
+                prompt,
+            )
+
+        return commands
+
+    @staticmethod
+    def _filter_redundant_start_flow_commands(
+        tracker: DialogueStateTracker, commands: List[Command]
+    ) -> List[Command]:
+        """Filters out StartFlowCommand commands for flows that are already active,
+        based on the current tracker state.
+        """
+        frames = tracker.stack.frames
+        active_user_flows = {
+            frame.flow_id for frame in frames if isinstance(frame, UserFlowStackFrame)
+        }
+        commands = [
+            command
+            for command in commands
+            if not (
+                isinstance(command, StartFlowCommand)
+                and command.flow in active_user_flows
+            )
+        ]
+        return commands
+
+    async def _predict_commands_for_newly_started_flows(
+        self,
+        message: Message,
+        tracker: DialogueStateTracker,
+        newly_started_flows: FlowsList,
+        all_flows: FlowsList,
+    ) -> List[Command]:
+        """Predict set slot commands for newly started flows."""
+        commands_for_newly_started_flows = []
+        for newly_started_flow in newly_started_flows:
+            commands_for_newly_started_flows += (
+                await self._predict_commands_for_newly_started_flow(
+                    newly_started_flow, message, tracker, newly_started_flows
+                )
+            )
+        return commands_for_newly_started_flows
+
+    async def _predict_commands_for_newly_started_flow(
+        self,
+        newly_started_flow: Flow,
+        message: Message,
+        tracker: DialogueStateTracker,
+        newly_started_flows: FlowsList,
+    ) -> List[Command]:
+        inputs = self._prepare_inputs_for_single_flow(
+            message, tracker, newly_started_flow, max_turns=20
+        )
+
+        if len(inputs["flow_slots"]) == 0:
+            # return empty if the newly started flow does not have any slots
+            return []
+
+        prompt = Template(self.fill_slots_prompt).render(**inputs)
+        structlogger.debug(
+            "multi_step_llm_command_generator"
+            ".predict_commands_for_newly_started_flow"
+            ".prompt_rendered",
+            flow=newly_started_flow.id,
+            prompt=prompt,
+        )
+
+        actions = await self.invoke_llm(prompt)
+        structlogger.debug(
+            "multi_step_llm_command_generator"
+            ".predict_commands_for_newly_started_flow"
+            ".actions_generated",
+            flow=newly_started_flow.id,
+            action_list=actions,
+        )
+
+        commands = self.parse_commands(actions, tracker, newly_started_flows)
+
+        # filter out all commands that unset values for newly started flow
+        commands = [
+            command
+            for command in commands
+            if isinstance(command, SetSlotCommand) and command.value
+        ]
+        structlogger.debug(
+            "multi_step_llm_command_generator"
+            ".predict_commands_for_newly_started_flow"
+            ".filtered_commands",
+            flow=newly_started_flow.id,
+            commands=commands,
+        )
+
+        if commands:
+            self._add_prompt_to_message_parse_data(
+                message,
+                MultiStepLLMCommandGenerator.__name__,
+                "fill_slots_for_new_flow_prompt",
+                prompt,
+            )
+
+        return commands
+
+    def _prepare_inputs(
+        self,
+        message: Message,
+        tracker: DialogueStateTracker,
+        available_flows: FlowsList,
+        all_flows: FlowsList,
+        max_turns: int = 1,
+    ) -> Dict[str, Any]:
+        """Prepare input information to be used by prompt template.
+
+        Args:
+            message: The message from the user.
+            tracker: The tracker containing the current state of the conversation.
+            available_flows: Startable and active flows.
+            all_flows: All flows.
+            max_turns: Max turns of the conversation history between the user
+                and the assistant
+
+        Returns:
+            Dictionary of inputs.
+        """
+        top_relevant_frame = top_flow_frame(tracker.stack)
+        top_flow = top_relevant_frame.flow(all_flows) if top_relevant_frame else None
+        current_step = (
+            top_relevant_frame.step(all_flows) if top_relevant_frame else None
+        )
+        if top_flow is not None:
+            flow_slots = self.prepare_current_flow_slots_for_template(
+                top_flow, current_step, tracker
+            )
+            top_flow_is_pattern = top_flow.is_rasa_default_flow
+        else:
+            flow_slots = []
+            top_flow_is_pattern = False
+
+        if top_flow_is_pattern:
+            top_user_frame = top_user_flow_frame(tracker.stack)
+            top_user_flow = (
+                top_user_frame.flow(available_flows) if top_user_frame else None
+            )
+            top_user_flow_step = (
+                top_user_frame.step(available_flows) if top_user_frame else None
+            )
+            top_user_flow_slots = self.prepare_current_flow_slots_for_template(
+                top_user_flow, top_user_flow_step, tracker
+            )
+        else:
+            top_user_flow = None
+            top_user_flow_slots = []
+
+        current_slot, current_slot_description = self.prepare_current_slot_for_template(
+            current_step
+        )
+        current_slot_type = None
+        current_slot_allowed_values = None
+        if current_slot:
+            current_slot_type = (
+                slot.type_name
+                if (slot := tracker.slots.get(current_slot)) is not None
+                else None
+            )
+            current_slot_allowed_values = allowed_values_for_slot(
+                tracker.slots.get(current_slot)
+            )
+        (
+            current_conversation,
+            latest_user_message,
+        ) = self.prepare_conversation_context_for_template(message, tracker, max_turns)
+
+        inputs = {
+            "available_flows": self.prepare_flows_for_template(
+                available_flows, tracker
+            ),
+            "current_conversation": current_conversation,
+            "current_flow": top_flow.id if top_flow is not None else None,
+            "current_slot": current_slot,
+            "current_slot_description": current_slot_description,
+            "current_slot_type": current_slot_type,
+            "current_slot_allowed_values": current_slot_allowed_values,
+            "last_user_message": latest_user_message,
+            "flow_slots": flow_slots,
+            "top_flow_is_pattern": top_flow_is_pattern,
+            "top_user_flow": top_user_flow.id if top_user_flow is not None else None,
+            "top_user_flow_slots": top_user_flow_slots,
+            "flow_active": True,
+        }
+        return inputs
+
+    def _prepare_inputs_for_single_flow(
+        self,
+        message: Message,
+        tracker: DialogueStateTracker,
+        flow: Flow,
+        max_turns: int = 1,
+    ) -> Dict[Text, Any]:
+        flow_slots = self.prepare_current_flow_slots_for_template(
+            flow, flow.first_step_in_flow(), tracker
+        )
+        (
+            current_conversation,
+            latest_user_message,
+        ) = self.prepare_conversation_context_for_template(message, tracker, max_turns)
+        inputs = {
+            "current_conversation": current_conversation,
+            "flow_slots": flow_slots,
+            "current_flow": flow.id,
+            "last_user_message": latest_user_message,
+            "flow_active": False,
+        }
+        return inputs
+
+    @classmethod
+    def fingerprint_addon(cls, config: Dict[str, Any]) -> Optional[str]:
+        """Add a fingerprint for the graph."""
+        handle_flows_template = get_prompt_template(
+            config.get("prompt_templates", {})
+            .get(HANDLE_FLOWS_KEY, {})
+            .get(FILE_PATH_KEY),
+            DEFAULT_HANDLE_FLOWS_TEMPLATE,
+        )
+        fill_slots_template = get_prompt_template(
+            config.get("prompt_templates", {})
+            .get(FILL_SLOTS_KEY, {})
+            .get(FILE_PATH_KEY),
+            DEFAULT_FILL_SLOTS_TEMPLATE,
+        )
+
+        llm_config = resolve_model_client_config(
+            config.get(LLM_CONFIG_KEY), MultiStepLLMCommandGenerator.__name__
+        )
+        embedding_config = resolve_model_client_config(
+            config.get(FLOW_RETRIEVAL_KEY, {}).get(EMBEDDINGS_CONFIG_KEY),
+            FlowRetrieval.__name__,
+        )
+
+        return deep_container_fingerprint(
+            [handle_flows_template, fill_slots_template, llm_config, embedding_config]
+        )
+
+    @staticmethod
+    def prepare_conversation_context_for_template(
+        message: Message, tracker: DialogueStateTracker, max_turns: int = 20
+    ) -> Tuple[Text, Text]:
+        current_conversation = tracker_as_readable_transcript(
+            tracker, max_turns=max_turns
+        )
+        latest_user_message = sanitize_message_for_prompt(message.get(TEXT))
+        current_conversation += f"\nUSER: {latest_user_message}"
+        return current_conversation, latest_user_message
+
+    def prepare_current_flow_slots_for_template(
+        self, top_flow: Flow, current_step: FlowStep, tracker: DialogueStateTracker
+    ) -> List[Dict[Text, Any]]:
+        """Prepare the current flow slots for the template.
+
+        Args:
+            top_flow: The top flow.
+            current_step: The current step in the flow.
+            tracker: The tracker containing the current state of the conversation.
+
+        Returns:
+            The slots with values, types, allowed values and a description.
+        """
+        if top_flow is not None:
+            flow_slots = [
+                {
+                    "name": collect_step.collect,
+                    "value": self.get_slot_value(tracker, collect_step.collect),
+                    "type": tracker.slots[collect_step.collect].type_name,
+                    "allowed_values": allowed_values_for_slot(
+                        tracker.slots[collect_step.collect]
+                    ),
+                    "description": collect_step.description,
+                }
+                for collect_step in top_flow.get_collect_steps()
+                if self.is_extractable(collect_step, tracker, current_step)
+            ]
+        else:
+            flow_slots = []
+        return flow_slots
+
+    @staticmethod
+    def prepare_current_slot_for_template(
+        current_step: FlowStep,
+    ) -> Tuple[Union[str, None], Union[str, None]]:
+        """Prepare the current slot for the template."""
+        return (
+            (current_step.collect, current_step.description)
+            if isinstance(current_step, CollectInformationFlowStep)
+            else (None, None)
+        )
+
+    @staticmethod
+    def _clean_up_commands(commands: List[Command]) -> List[Command]:
+        """Cleans the list of commands by removing CannotHandleCommand,
+        if it exists and there are other commands in the list.
+        """
+        other_commands_count = sum(
+            not isinstance(command, CannotHandleCommand) for command in commands
+        )
+
+        if other_commands_count == len(commands):
+            # no cannot handle command found
+            return commands
+
+        if other_commands_count:
+            # remove cannot handle commands
+            return [
+                command
+                for command in commands
+                if not isinstance(command, CannotHandleCommand)
+            ]
+
+        # only cannot handle commands present
+        return [CannotHandleCommand(RASA_PATTERN_CANNOT_HANDLE_NOT_SUPPORTED)]