rasa-pro 3.12.0.dev13__py3-none-any.whl → 3.12.0rc2__py3-none-any.whl

rasa/dialogue_understanding/generator/single_step/compact_llm_command_generator.py

@@ -0,0 +1,522 @@
+from typing import Any, Dict, List, Optional, Text
+
+import structlog
+
+import rasa.shared.utils.io
+from rasa.dialogue_understanding.commands import (
+    CannotHandleCommand,
+    Command,
+    ErrorCommand,
+    SetSlotCommand,
+)
+from rasa.dialogue_understanding.commands.command_syntax_manager import (
+    CommandSyntaxManager,
+    CommandSyntaxVersion,
+)
+from rasa.dialogue_understanding.generator import LLMBasedCommandGenerator
+from rasa.dialogue_understanding.generator.command_parser import (
+    parse_commands as parse_commands_using_command_parsers,
+)
+from rasa.dialogue_understanding.generator.constants import (
+    COMMAND_PROMPT_FILE_NAME,
+    DEFAULT_LLM_CONFIG,
+    FLOW_RETRIEVAL_KEY,
+    LLM_BASED_COMMAND_GENERATOR_CONFIG_FILE,
+    LLM_CONFIG_KEY,
+    USER_INPUT_CONFIG_KEY,
+)
+from rasa.dialogue_understanding.generator.flow_retrieval import FlowRetrieval
+from rasa.dialogue_understanding.stack.utils import top_flow_frame
+from rasa.dialogue_understanding.utils import (
+    add_commands_to_message_parse_data,
+    add_prompt_to_message_parse_data,
+)
+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 (
+    EMBEDDINGS_CONFIG_KEY,
+    PROMPT_TEMPLATE_CONFIG_KEY,
+    ROUTE_TO_CALM_SLOT,
+)
+from rasa.shared.core.flows import FlowsList
+from rasa.shared.core.trackers import DialogueStateTracker
+from rasa.shared.exceptions import ProviderClientAPIException
+from rasa.shared.nlu.constants import LLM_COMMANDS, LLM_PROMPT, TEXT
+from rasa.shared.nlu.training_data.message import Message
+from rasa.shared.providers.llm.llm_response import LLMResponse
+from rasa.shared.utils.io import deep_container_fingerprint
+from rasa.shared.utils.llm import (
+    allowed_values_for_slot,
+    get_default_prompt_template_based_on_model,
+    get_prompt_template,
+    resolve_model_client_config,
+    sanitize_message_for_prompt,
+    tracker_as_readable_transcript,
+)
+from rasa.utils.beta import BetaNotEnabledException, ensure_beta_feature_is_enabled
+from rasa.utils.log_utils import log_llm
+
+structlogger = structlog.get_logger()
+
+
+MODEL_PROMPT_MAPPER = {
+    "openai/gpt-4o-2024-11-20": "command_prompt_v2_gpt_4o_2024_11_20_template.jinja2",
+    "azure/gpt-4o-2024-11-20": "command_prompt_v2_gpt_4o_2024_11_20_template.jinja2",
+    "bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0": (
+        "command_prompt_v2_claude_3_5_sonnet_20240620_template.jinja2"
+    ),
+    "anthropic/claude-3-5-sonnet-20240620": (
+        "command_prompt_v2_claude_3_5_sonnet_20240620_template.jinja2"
+    ),
+}
+
+DEFAULT_COMMAND_PROMPT_TEMPLATE_FILE_NAME = "command_prompt_v2_default.jinja2"
+
+
+class CommandParserValidatorSingleton:
+    """Singleton class to validate the command parser.
+
+    This class is used to validate the command parser. It keeps track of the number of
+    consecutive turns where no commands are parsed by the command parser. If the
+    number of consecutive turns exceeds a certain threshold, a warning is logged.
+    The prompt can use a DSL syntax that can be incompatible with the command syntax
+    used by the command parser. This class helps to detect such incompatibilities.
+    """
+
+    MAX_CONSECUTIVE_TURNS_NO_COMMAND_PREDICTED = 5
+    _NO_COMMAND_PREDICTED_TURN_COUNTER = 0
+    _command_parser_validated = False
+
+    @classmethod
+    def get_no_command_predicted_turn_counter(cls) -> int:
+        return cls._NO_COMMAND_PREDICTED_TURN_COUNTER
+
+    @classmethod
+    def should_validate_command_parser(cls) -> bool:
+        return not cls._command_parser_validated
+
+    @classmethod
+    def reset_command_parser_validation(cls) -> None:
+        cls._NO_COMMAND_PREDICTED_TURN_COUNTER = 0
+        cls._command_parser_validated = False
+
+    @classmethod
+    def validate_if_commands_are_parsed_from_llm_response(
+        cls, commands: List[Command], llm_response: str
+    ) -> None:
+        if llm_response and not commands:
+            cls._NO_COMMAND_PREDICTED_TURN_COUNTER += 1
+        else:
+            # Reset the counter if commands are generated, and mark
+            # the command parser as validated.
+            cls._NO_COMMAND_PREDICTED_TURN_COUNTER = 0
+            cls._command_parser_validated = True
+            return
+
+        if (
+            cls._NO_COMMAND_PREDICTED_TURN_COUNTER
+            >= cls.MAX_CONSECUTIVE_TURNS_NO_COMMAND_PREDICTED
+        ):
+            structlogger.warning(
+                "llm_command_generator.predict_commands.command_parser_not_working",
+                event_info=(
+                    f"No commands were generated by the command parser for the last "
+                    f"{cls._NO_COMMAND_PREDICTED_TURN_COUNTER} times. Check if you "
+                    "are running incompatible prompt and LLM command generator."
+                ),
+            )
+
+
+@DefaultV1Recipe.register(
+    [
+        DefaultV1Recipe.ComponentType.COMMAND_GENERATOR,
+    ],
+    is_trainable=True,
+)
+class CompactLLMCommandGenerator(LLMBasedCommandGenerator):
+    """A single step LLM-based command generator."""
+
+    def __init__(
+        self,
+        config: Dict[str, Any],
+        model_storage: ModelStorage,
+        resource: Resource,
+        prompt_template: Optional[Text] = None,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(
+            config,
+            model_storage,
+            resource,
+            prompt_template=prompt_template,
+            **kwargs,
+        )
+
+        # Get the default prompt template based on the model name
+        default_command_prompt_template = get_default_prompt_template_based_on_model(
+            self.config.get(LLM_CONFIG_KEY, {}) or {},
+            MODEL_PROMPT_MAPPER,
+            DEFAULT_COMMAND_PROMPT_TEMPLATE_FILE_NAME,
+        )
+
+        # Set the prompt template either from the config or the default prompt template.
+        self.prompt_template = prompt_template or get_prompt_template(
+            self.config.get(PROMPT_TEMPLATE_CONFIG_KEY),
+            default_command_prompt_template,
+        )
+
+        self.trace_prompt_tokens = self.config.get("trace_prompt_tokens", False)
+        self.repeat_command_enabled = self.is_repeat_command_enabled()
+
+        # Set the command syntax version to v2
+        CommandSyntaxManager.set_syntax_version(CommandSyntaxVersion.v2)
+
+    ### 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_TEMPLATE_CONFIG_KEY: None,
+            USER_INPUT_CONFIG_KEY: None,
+            LLM_CONFIG_KEY: None,
+            FLOW_RETRIEVAL_KEY: FlowRetrieval.get_default_config(),
+        }
+
+    def persist(self) -> None:
+        """Persist this component to disk for future loading."""
+        self._persist_prompt_template()
+        self._persist_config()
+        if self.flow_retrieval is not None:
+            self.flow_retrieval.persist()
+
+    def _persist_prompt_template(self) -> None:
+        """Persist prompt template for future loading."""
+        with self._model_storage.write_to(self._resource) as path:
+            rasa.shared.utils.io.write_text_file(
+                self.prompt_template, path / COMMAND_PROMPT_FILE_NAME
+            )
+
+    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 / LLM_BASED_COMMAND_GENERATOR_CONFIG_FILE, self.config
+            )
+
+    @classmethod
+    def load(
+        cls: Any,
+        config: Dict[str, Any],
+        model_storage: ModelStorage,
+        resource: Resource,
+        execution_context: ExecutionContext,
+        **kwargs: Any,
+    ) -> "CompactLLMCommandGenerator":
+        """Loads trained component (see parent class for full docstring)."""
+        # Perform health check of the LLM API endpoint
+        llm_config = resolve_model_client_config(config.get(LLM_CONFIG_KEY, {}))
+        cls.perform_llm_health_check(
+            llm_config,
+            DEFAULT_LLM_CONFIG,
+            "compact_llm_command_generator.load",
+            cls.__name__,
+        )
+
+        # load prompt template from the model storage.
+        prompt_template = cls.load_prompt_template_from_model_storage(
+            model_storage, resource, COMMAND_PROMPT_FILE_NAME
+        )
+
+        # init base command generator
+        command_generator = cls(config, model_storage, resource, prompt_template)
+        # 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
+
+    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.
+        """
+        prior_commands = self._get_prior_commands(message)
+
+        if tracker is None or flows.is_empty():
+            # cannot do anything if there are no flows or no tracker
+            return prior_commands
+
+        if self._should_skip_llm_call(prior_commands, flows, tracker):
+            return prior_commands
+
+        try:
+            commands = await self._predict_commands(message, flows, tracker)
+        except ProviderClientAPIException:
+            # if command predictions resulted in API exception
+            # "predict" the ErrorCommand
+            commands = [ErrorCommand()]
+            structlogger.warning(
+                "llm_command_generator.predict_commands.api_exception",
+                event_info=(
+                    "ProviderClientAPIException occurred while predicting commands."
+                ),
+                commands=commands,
+            )
+
+        if not commands and not prior_commands:
+            # no commands are parsed or there's an invalid command
+            structlogger.warning(
+                "llm_command_generator.predict_commands",
+                message="No commands were predicted as the LLM response could "
+                "not be parsed or the LLM responded with an invalid command. "
+                "Returning a CannotHandleCommand instead.",
+            )
+            commands = [CannotHandleCommand()]
+
+        if tracker.has_coexistence_routing_slot:
+            # if coexistence feature is used, set the routing slot
+            commands += [SetSlotCommand(ROUTE_TO_CALM_SLOT, True)]
+
+        log_llm(
+            logger=structlogger,
+            log_module=self.__class__.__name__,
+            log_event="llm_command_generator.predict_commands.finished",
+            commands=commands,
+        )
+
+        domain = kwargs.get("domain")
+        commands = self._check_commands_against_slot_mappings(commands, tracker, domain)
+
+        return self._check_commands_overlap(prior_commands, commands)
+
+    async def _predict_commands(
+        self,
+        message: Message,
+        flows: FlowsList,
+        tracker: Optional[DialogueStateTracker] = None,
+    ) -> 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 flows
+        filtered_flows = await self.filter_flows(message, flows, tracker)
+
+        flow_prompt = self.render_template(message, tracker, filtered_flows, flows)
+        log_llm(
+            logger=structlogger,
+            log_module=self.__class__.__name__,
+            log_event="llm_command_generator.predict_commands.prompt_rendered",
+            prompt=flow_prompt,
+        )
+
+        response = await self.invoke_llm(flow_prompt)
+        llm_response = LLMResponse.ensure_llm_response(response)
+        # The check for 'None' maintains compatibility with older versions
+        # of LLMCommandGenerator. In previous implementations, 'invoke_llm'
+        # might return 'None' to indicate a failure to generate actions.
+        if llm_response is None or not llm_response.choices:
+            structlogger.warning(
+                "llm_command_generator.predict_commands.no_actions_generated",
+                event_info=(
+                    "No actions were generated by the LLM. Returning an ErrorCommand."
+                ),
+            )
+            return [ErrorCommand()]
+
+        action_list = llm_response.choices[0]
+
+        log_llm(
+            logger=structlogger,
+            log_module=self.__class__.__name__,
+            log_event="llm_command_generator.predict_commands.actions_generated",
+            action_list=action_list,
+        )
+
+        commands = self.parse_commands(action_list, tracker, flows)
+
+        if CommandParserValidatorSingleton.should_validate_command_parser():
+            CommandParserValidatorSingleton.validate_if_commands_are_parsed_from_llm_response(
+                commands, action_list
+            )
+
+        self._update_message_parse_data_for_fine_tuning(message, commands, flow_prompt)
+        add_commands_to_message_parse_data(message, self.__class__.__name__, commands)
+        add_prompt_to_message_parse_data(
+            message=message,
+            component_name=self.__class__.__name__,
+            prompt_name="command_generator_prompt",
+            user_prompt=flow_prompt,
+            llm_response=llm_response,
+        )
+
+        return commands
+
+    @staticmethod
+    def _update_message_parse_data_for_fine_tuning(
+        message: Message, commands: List[Command], prompt: str
+    ) -> None:
+        from rasa.llm_fine_tuning.annotation_module import preparing_fine_tuning_data
+
+        if preparing_fine_tuning_data:
+            # Add commands and prompt to the message object in order to create
+            # prompt -> commands pairs for fine-tuning
+            message.set(
+                LLM_COMMANDS,
+                [command.as_dict() for command in commands],
+                add_to_output=True,
+            )
+            message.set(LLM_PROMPT, prompt, add_to_output=True)
+
+    @classmethod
+    def parse_commands(
+        cls, actions: Optional[str], tracker: DialogueStateTracker, flows: FlowsList
+    ) -> 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
+
+        Returns:
+            The parsed commands.
+        """
+        commands = parse_commands_using_command_parsers(actions, flows)
+        if not commands:
+            structlogger.warning(
+                f"{cls.__name__}.parse_commands",
+                message="No commands were parsed from the LLM actions.",
+                actions=actions,
+            )
+
+        return commands
+
+    ### Helper methods
+    def render_template(
+        self,
+        message: Message,
+        tracker: DialogueStateTracker,
+        startable_flows: FlowsList,
+        all_flows: FlowsList,
+    ) -> str:
+        """Render the jinja template to create the prompt for the LLM.
+
+        Args:
+            message: The current message from the user.
+            tracker: The tracker containing the current state of the conversation.
+            startable_flows: The flows startable at this point in time by the user.
+            all_flows: all flows present in the assistant
+
+        Returns:
+            The rendered prompt template.
+        """
+        # need to make this distinction here because current step of the
+        # top_calling_frame would be the call step, but we need the collect step from
+        # the called frame. If no call is active calling and called frame are the same.
+        top_calling_frame = top_flow_frame(tracker.stack)
+        top_called_frame = top_flow_frame(tracker.stack, ignore_call_frames=False)
+
+        top_flow = top_calling_frame.flow(all_flows) if top_calling_frame else None
+        current_step = top_called_frame.step(all_flows) if top_called_frame else None
+
+        flow_slots = self.prepare_current_flow_slots_for_template(
+            top_flow, current_step, tracker
+        )
+        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 = tracker_as_readable_transcript(tracker)
+        latest_user_message = sanitize_message_for_prompt(message.get(TEXT))
+        current_conversation += f"\nUSER: {latest_user_message}"
+
+        inputs = {
+            "available_flows": self.prepare_flows_for_template(
+                startable_flows, tracker
+            ),
+            "current_conversation": current_conversation,
+            "flow_slots": flow_slots,
+            "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,
+            "user_message": latest_user_message,
+            "is_repeat_command_enabled": self.repeat_command_enabled,
+        }
+
+        return self.compile_template(self.prompt_template).render(**inputs)
+
+    def is_repeat_command_enabled(self) -> bool:
+        """Check for feature flag"""
+        RASA_PRO_BETA_REPEAT_COMMAND_ENV_VAR_NAME = "RASA_PRO_BETA_REPEAT_COMMAND"
+        try:
+            ensure_beta_feature_is_enabled(
+                "Repeat Command",
+                env_flag=RASA_PRO_BETA_REPEAT_COMMAND_ENV_VAR_NAME,
+            )
+        except BetaNotEnabledException:
+            return False
+
+        return True
+
+    @classmethod
+    def fingerprint_addon(cls: Any, config: Dict[str, Any]) -> Optional[str]:
+        """Add a fingerprint for the graph."""
+        # Get the default prompt template based on the model name
+        llm_config = resolve_model_client_config(
+            config.get(LLM_CONFIG_KEY), CompactLLMCommandGenerator.__name__
+        )
+        embedding_config = resolve_model_client_config(
+            config.get(FLOW_RETRIEVAL_KEY, {}).get(EMBEDDINGS_CONFIG_KEY),
+            FlowRetrieval.__name__,
+        )
+        default_command_prompt_template = get_default_prompt_template_based_on_model(
+            llm_config or {},
+            MODEL_PROMPT_MAPPER,
+            DEFAULT_COMMAND_PROMPT_TEMPLATE_FILE_NAME,
+        )
+        prompt_template = get_prompt_template(
+            config.get(PROMPT_TEMPLATE_CONFIG_KEY),
+            default_command_prompt_template,
+        )
+        return deep_container_fingerprint(
+            [prompt_template, llm_config, embedding_config]
+        )