rasa-pro 3.12.0rc1__py3-none-any.whl → 3.12.0rc3__py3-none-any.whl

README.md

@@ -2,15 +2,14 @@
 
 <div align="center">
 
-[![Build Status](https://github.com/RasaHQ/rasa-private/workflows/Continuous%20Integration/badge.svg)](https://github.com/RasaHQ/rasa-private/actions)
 [![Quality Gate Status](https://sonarcloud.io/api/project_badges/measure?project=RasaHQ_rasa&metric=alert_status)](https://sonarcloud.io/summary/new_code?id=RasaHQ_rasa)
-[![Documentation Status](https://img.shields.io/badge/docs-stable-brightgreen.svg)](https://rasa.com/docs/rasa-pro/)
+[![Documentation Status](https://img.shields.io/badge/docs-stable-brightgreen.svg)](https://rasa.com/docs/docs/pro/intro)
+![Python version support](https://img.shields.io/pypi/pyversions/rasa-pro)
 
 </div>
 
 <hr />
 
-
 Rasa Pro is a framework for building scalable, dynamic conversational AI assistants that integrate large language models (LLMs) to enable more contextually aware and agentic interactions. Whether you’re new to conversational AI or an experienced developer, Rasa Pro offers enhanced flexibility, control, and performance for mission-critical applications.
 
 Building on the foundation of Rasa Open Source, Rasa Pro adds advanced features like CALM (Conversational AI with Language Models) and Dialogue Understanding (DU), which enable developers to shift from traditional intent-driven systems to LLM-based agents. This allows for more robust, responsive interactions that adhere strictly to business logic, while reducing risks like prompt injection and minimizing hallucinations.
@@ -23,19 +22,17 @@ Building on the foundation of Rasa Open Source, Rasa Pro adds advanced features
 - **Robustness and Control:** Maintain strict adherence to business logic, preventing unwanted behaviors like prompt injection and hallucinations, leading to more reliable responses and secure interactions.
 - **Built-in Security:** Safeguard sensitive data, control access, and ensure secure deployment, essential for production environments that demand high levels of security and compliance.
 
+A [free developer license](https://rasa.com/docs/pro/intro/#who-rasa-pro-is-for) is available so you can explore and get to know Rasa Pro. It allows you to take your assistant live in production a limited capacity. A paid license is required for larger-scale production use, but all code is visible and can be customized as needed.
 
+To get started right now, you can
 
-A [free developer license](https://rasa.com/docs/rasa-pro/developer-edition/) is available so you can explore and get to know Rasa Pro. For small production deployments, the Extended Developer License allows you to take your assistant live in a limited capacity. A paid license is required for larger-scale production use, but all code is visible and can be customized as needed.
-
-To get started right now, you can 
-
-`pip install rasa-pro` 
+`pip install rasa-pro`
 
-Check out our 
+Check out our
 
-- [Rasa-pro Quickstart](https://rasa.com/docs/rasa-pro/installation/quickstart/),
-- [Conversational AI with Language Models (CALM) conceptual rundown](https://rasa.com/docs/rasa-pro/calm/),
-- [Rasa Pro / CALM tutorial](https://rasa.com/docs/rasa-pro/tutorial), and
-- [Rasa pro changelog](https://rasa.com/docs/rasa/rasa-pro-changelog/)
+- [Rasa-pro Quickstart](https://rasa.com/docs/learn/quickstart/pro),
+- [Conversational AI with Language Models (CALM) conceptual rundown](https://rasa.com/docs/learn/concepts/calm),
+- [Rasa Pro / CALM tutorial](https://rasa.com/docs/pro/tutorial), and
+- [Rasa pro changelog](https://rasa.com/docs/reference/changelogs/rasa-pro-changelog)
 
 for more. Also feel free to reach out to us on the [Rasa forum](https://forum.rasa.com/).

rasa/cli/dialogue_understanding_test.py

@@ -3,7 +3,7 @@ import asyncio
 import datetime
 import importlib
 import sys
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, Type, cast
 
 import structlog
 
@@ -20,9 +20,7 @@ from rasa.core.exceptions import AgentNotReady
 from rasa.core.processor import MessageProcessor
 from rasa.core.utils import AvailableEndpoints
 from rasa.dialogue_understanding.commands import Command
-from rasa.dialogue_understanding.generator import (
-    LLMBasedCommandGenerator,
-)
+from rasa.dialogue_understanding.generator import LLMBasedCommandGenerator
 from rasa.dialogue_understanding.generator.command_parser import DEFAULT_COMMANDS
 from rasa.dialogue_understanding_test.command_metric_calculation import (
     calculate_command_metrics,
@@ -372,18 +370,17 @@ def split_test_results(
 def _get_llm_command_generator_config(
     processor: MessageProcessor,
 ) -> Optional[Dict[str, Any]]:
-    from rasa.dialogue_understanding.generator.constants import DEFAULT_LLM_CONFIG
-
     train_schema = processor.model_metadata.train_schema
 
     for node_name, node in train_schema.nodes.items():
         if node.matches_type(LLMBasedCommandGenerator, include_subtypes=True):
             # Configurations can reference model groups defined in the endpoints.yml
-            resolved_config = resolve_model_client_config(
+            resolved_llm_config = resolve_model_client_config(
                 node.config.get(LLM_CONFIG_KEY, {}), node_name
             )
+            llm_command_generator = cast(Type[LLMBasedCommandGenerator], node.uses)
             return combine_custom_and_default_config(
-                resolved_config, DEFAULT_LLM_CONFIG
+                resolved_llm_config, llm_command_generator.get_default_llm_config()
             )
 
     return None

rasa/cli/llm_fine_tuning.py

@@ -1,7 +1,7 @@
 import argparse
 import asyncio
 import sys
-from typing import Any, Dict, List
+from typing import Any, Dict, List, Type, cast
 
 import structlog
 
@@ -22,7 +22,12 @@ from rasa.cli.e2e_test import (
 )
 from rasa.core.exceptions import AgentNotReady
 from rasa.core.utils import AvailableEndpoints
-from rasa.dialogue_understanding.generator import SingleStepLLMCommandGenerator
+from rasa.dialogue_understanding.generator.llm_based_command_generator import (
+    LLMBasedCommandGenerator,
+)
+from rasa.dialogue_understanding.generator.multi_step.multi_step_llm_command_generator import (  # noqa: E501
+    MultiStepLLMCommandGenerator,
+)
 from rasa.e2e_test.e2e_test_runner import E2ETestRunner
 from rasa.llm_fine_tuning.annotation_module import annotate_e2e_tests
 from rasa.llm_fine_tuning.llm_data_preparation_module import convert_to_fine_tuning_data
@@ -112,7 +117,6 @@ def create_llm_finetune_data_preparation_subparser(
         help_text="Configuration file for the model server and the connectors as a "
         "yml file.",
     )
-
     return data_preparation_subparser
 
 
@@ -205,6 +209,9 @@ def prepare_llm_fine_tuning_data(args: argparse.Namespace) -> None:
 
     flows = asyncio.run(e2e_test_runner.agent.processor.get_flows())
     llm_command_generator_config = _get_llm_command_generator_config(e2e_test_runner)
+    llm_command_generator: Type[LLMBasedCommandGenerator] = _get_llm_command_generator(
+        e2e_test_runner
+    )
 
     # set up storage context
     storage_context = create_storage_context(StorageType.FILE, output_dir)
@@ -235,6 +242,7 @@ def prepare_llm_fine_tuning_data(args: argparse.Namespace) -> None:
             rephrase_config,
             args.num_rephrases,
             flows,
+            llm_command_generator,
             llm_command_generator_config,
             storage_context,
         )
@@ -271,30 +279,57 @@ def prepare_llm_fine_tuning_data(args: argparse.Namespace) -> None:
     write_statistics(statistics, output_dir)
 
     rasa.shared.utils.cli.print_success(
-        f"Data and intermediate results are written " f"to '{output_dir}'."
+        f"Data and intermediate results are written to '{output_dir}'."
     )
 
 
 def _get_llm_command_generator_config(e2e_test_runner: E2ETestRunner) -> Dict[str, Any]:
-    from rasa.dialogue_understanding.generator.constants import DEFAULT_LLM_CONFIG
-
     train_schema = e2e_test_runner.agent.processor.model_metadata.train_schema  # type: ignore
 
     for node_name, node in train_schema.nodes.items():
-        if node.matches_type(SingleStepLLMCommandGenerator, include_subtypes=True):
+        if node.matches_type(
+            LLMBasedCommandGenerator, include_subtypes=True
+        ) and not node.matches_type(
+            MultiStepLLMCommandGenerator, include_subtypes=True
+        ):
             # Configurations can reference model groups defined in the endpoints.yml
-            resolved_config = resolve_model_client_config(
+            resolved_llm_config = resolve_model_client_config(
                 node.config.get(LLM_CONFIG_KEY, {}), node_name
             )
+            llm_command_generator = cast(Type[LLMBasedCommandGenerator], node.uses)
             return combine_custom_and_default_config(
-                resolved_config, DEFAULT_LLM_CONFIG
+                resolved_llm_config, llm_command_generator.get_default_llm_config()
             )
 
     rasa.shared.utils.cli.print_error(
         "The provided model is not trained using 'SingleStepLLMCommandGenerator' or "
-        "its subclasses. Without it, no data for fine-tuning can be generated. To "
-        "resolve this, please include 'SingleStepLLMCommandGenerator' or its subclass "
-        "in your config and train your model."
+        "'CompactLLMCommandGenerator' or its subclasses. Without it, no data for "
+        "fine-tuning can be generated. To resolve this, please include "
+        "'SingleStepLLMCommandGenerator' or 'CompactLLMCommandGenerator' or its "
+        "subclasses in your config and train your model."
+    )
+    sys.exit(1)
+
+
+def _get_llm_command_generator(
+    e2e_test_runner: E2ETestRunner,
+) -> Type[LLMBasedCommandGenerator]:
+    train_schema = e2e_test_runner.agent.processor.model_metadata.train_schema  # type: ignore
+
+    for _, node in train_schema.nodes.items():
+        if node.matches_type(
+            LLMBasedCommandGenerator, include_subtypes=True
+        ) and not node.matches_type(
+            MultiStepLLMCommandGenerator, include_subtypes=True
+        ):
+            return cast(Type[LLMBasedCommandGenerator], node.uses)
+
+    rasa.shared.utils.cli.print_error(
+        "The provided model is not trained using 'SingleStepLLMCommandGenerator' or "
+        "'CompactLLMCommandGenerator' or its subclasses. Without it, no data for "
+        "fine-tuning can be generated. To resolve this, please include "
+        "'SingleStepLLMCommandGenerator' or 'CompactLLMCommandGenerator' or its "
+        "subclasses in your config and train your model."
     )
     sys.exit(1)
 

rasa/cli/project_templates/calm/domain/list_contacts.yml

@@ -7,8 +7,7 @@ slots:
   contacts_list:
     type: text
     mappings:
-      - type: custom
-        action: list_contacts
+      - type: controlled
 
 responses:
   utter_no_contacts:

rasa/cli/project_templates/calm/domain/remove_contact.yml

@@ -7,8 +7,7 @@ slots:
   remove_contact_name:
     type: text
     mappings:
-      - type: custom
-        action: remove_contact
+      - type: controlled
   remove_contact_handle:
     type: text
     mappings:

rasa/cli/project_templates/calm/domain/shared.yml

@@ -4,7 +4,4 @@ slots:
   return_value:
     type: any
     mappings:
-      - type: custom
-        action: add_contact
-      - type: custom
-        action: remove_contact
+      - type: controlled

rasa/core/actions/action_handle_digressions.py

@@ -18,6 +18,10 @@ from rasa.dialogue_understanding.stack.frames.flow_stack_frame import (
     FlowStackFrameType,
     UserFlowStackFrame,
 )
+from rasa.dialogue_understanding.stack.utils import (
+    remove_digression_from_stack,
+    user_flows_on_the_stack,
+)
 from rasa.shared.core.constants import (
     ACTION_BLOCK_DIGRESSION,
     ACTION_CONTINUE_DIGRESSION,
@@ -55,16 +59,24 @@ class ActionBlockDigressions(Action):
         frame_type = FlowStackFrameType.REGULAR
 
         stack = tracker.stack
-        stack.push(
-            UserFlowStackFrame(flow_id=blocked_flow_id, frame_type=frame_type), 0
-        )
-        stack.push(
-            ContinueInterruptedPatternFlowStackFrame(
-                previous_flow_name=blocked_flow_id
-            ),
-            1,
-        )
-        events = tracker.create_stack_updated_events(stack)
+
+        if blocked_flow_id in user_flows_on_the_stack(stack):
+            structlogger.debug(
+                "action_block_digressions.already_blocked_flow",
+                blocked_flow_id=blocked_flow_id,
+            )
+            events = []
+        else:
+            stack.push(
+                UserFlowStackFrame(flow_id=blocked_flow_id, frame_type=frame_type), 0
+            )
+            stack.push(
+                ContinueInterruptedPatternFlowStackFrame(
+                    previous_flow_name=blocked_flow_id
+                ),
+                1,
+            )
+            events = tracker.create_stack_updated_events(stack)
 
         utterance = "utter_block_digressions"
         message = await nlg.generate(
@@ -109,10 +121,20 @@ class ActionContinueDigression(Action):
         if not isinstance(top_frame, HandleDigressionsPatternFlowStackFrame):
             return []
 
-        blocked_flow_id = top_frame.interrupting_flow_id
-        frame_type = FlowStackFrameType.INTERRUPT
+        interrupting_flow_id = top_frame.interrupting_flow_id
         stack = tracker.stack
-        stack.push(UserFlowStackFrame(flow_id=blocked_flow_id, frame_type=frame_type))
+
+        if interrupting_flow_id in user_flows_on_the_stack(stack):
+            structlogger.debug(
+                "action_continue_digression.interrupting_flow_id_already_on_the_stack",
+                interrupting_flow_id=interrupting_flow_id,
+            )
+            stack = remove_digression_from_stack(stack, interrupting_flow_id)
+
+        frame_type = FlowStackFrameType.INTERRUPT
+        stack.push(
+            UserFlowStackFrame(flow_id=interrupting_flow_id, frame_type=frame_type)
+        )
 
         events = [
             FlowInterrupted(

rasa/core/channels/voice_stream/asr/asr_event.py

@@ -16,3 +16,8 @@ class NewTranscript(ASREvent):
 @dataclass
 class UserIsSpeaking(ASREvent):
     pass
+
+
+@dataclass
+class UserSilence(ASREvent):
+    pass

rasa/core/channels/voice_stream/audiocodes.py

@@ -46,6 +46,19 @@ class AudiocodesVoiceOutputChannel(VoiceOutputChannel):
     def name(cls) -> str:
         return "ac_voice"
 
+    def _ensure_stream_id(self) -> None:
+        """Audiocodes requires a stream ID with playStream messages."""
+        if "stream_id" not in call_state.channel_data:
+            call_state.channel_data["stream_id"] = 0
+
+    def _increment_stream_id(self) -> None:
+        self._ensure_stream_id()
+        call_state.channel_data["stream_id"] += 1
+
+    def _get_stream_id(self) -> str:
+        self._ensure_stream_id()
+        return str(call_state.channel_data["stream_id"])
+
     def rasa_audio_bytes_to_channel_bytes(
         self, rasa_audio_bytes: RasaAudioBytes
     ) -> bytes:
@@ -55,7 +68,7 @@ class AudiocodesVoiceOutputChannel(VoiceOutputChannel):
         media_message = json.dumps(
             {
                 "type": "playStream.chunk",
-                "streamId": str(call_state.stream_id),
+                "streamId": self._get_stream_id(),
                 "audioChunk": channel_bytes.decode("utf-8"),
             }
         )
@@ -63,14 +76,14 @@ class AudiocodesVoiceOutputChannel(VoiceOutputChannel):
 
     async def send_start_marker(self, recipient_id: str) -> None:
         """Send playStream.start before first audio chunk."""
-        call_state.stream_id += 1  # type: ignore[attr-defined]
+        self._increment_stream_id()
         media_message = json.dumps(
             {
                 "type": "playStream.start",
-                "streamId": str(call_state.stream_id),
+                "streamId": self._get_stream_id(),
             }
         )
-        logger.debug("Sending start marker", stream_id=call_state.stream_id)
+        logger.debug("Sending start marker", stream_id=self._get_stream_id())
         await self.voice_websocket.send(media_message)
 
     async def send_intermediate_marker(self, recipient_id: str) -> None:
@@ -82,10 +95,10 @@ class AudiocodesVoiceOutputChannel(VoiceOutputChannel):
         media_message = json.dumps(
             {
                 "type": "playStream.stop",
-                "streamId": str(call_state.stream_id),
+                "streamId": self._get_stream_id(),
             }
         )
-        logger.debug("Sending end marker", stream_id=call_state.stream_id)
+        logger.debug("Sending end marker", stream_id=self._get_stream_id())
         await self.voice_websocket.send(media_message)
 
 

rasa/core/channels/voice_stream/call_state.py

@@ -1,7 +1,7 @@
 import asyncio
 from contextvars import ContextVar
 from dataclasses import dataclass, field
-from typing import Optional
+from typing import Any, Dict, Optional
 
 from werkzeug.local import LocalProxy
 
@@ -19,14 +19,8 @@ class CallState:
     should_hangup: bool = False
     connection_failed: bool = False
 
-    # Genesys requires the server and client each maintain a
-    # monotonically increasing message sequence number.
-    client_sequence_number: int = 0
-    server_sequence_number: int = 0
-    audio_buffer: bytearray = field(default_factory=bytearray)
-
-    # Audiocodes requires a stream ID at start and end of stream
-    stream_id: int = 0
+    # Generic field for channel-specific state data
+    channel_data: Dict[str, Any] = field(default_factory=dict)
 
 
 _call_state: ContextVar[CallState] = ContextVar("call_state")

rasa/core/channels/voice_stream/genesys.py

@@ -27,8 +27,23 @@ from rasa.core.channels.voice_stream.voice_channel import (
     VoiceOutputChannel,
 )
 
-# Not mentioned in the documentation but observed in Geneys's example
-# https://github.com/GenesysCloudBlueprints/audioconnector-server-reference-implementation
+"""
+Genesys throws a rate limit error with too many audio messages.
+To avoid this, we buffer the audio messages and send them in chunks.
+
+- global.inbound.binary.average.rate.per.second: 5
+The allowed average rate per second of inbound binary data
+
+- global.inbound.binary.max: 25
+The maximum number of inbound binary data messages
+that can be sent instantaneously
+
+https://developer.genesys.cloud/organization/organization/limits#audiohook
+
+The maximum binary message size is not mentioned
+in the documentation but observed in their example app
+https://github.com/GenesysCloudBlueprints/audioconnector-server-reference-implementation
+"""
 MAXIMUM_BINARY_MESSAGE_SIZE = 64000  # 64KB
 logger = structlog.get_logger(__name__)
 
@@ -56,52 +71,7 @@ class GenesysOutputChannel(VoiceOutputChannel):
     async def send_audio_bytes(
         self, recipient_id: str, audio_bytes: RasaAudioBytes
     ) -> None:
-        """
-        Send audio bytes to the recipient with buffering.
-
-        Genesys throws a rate limit error with too many audio messages.
-        To avoid this, we buffer the audio messages and send them in chunks.
-
-        - global.inbound.binary.average.rate.per.second: 5
-        The allowed average rate per second of inbound binary data
-
-        - global.inbound.binary.max: 25
-        The maximum number of inbound binary data messages
-        that can be sent instantaneously
-
-        https://developer.genesys.cloud/organization/organization/limits#audiohook
-        """
-        call_state.audio_buffer.extend(audio_bytes)
-
-        # If we receive a non-standard chunk size, assume it's the end of a sequence
-        # or buffer is more than 32KB (this is half of genesys's max audio message size)
-        if len(audio_bytes) != 1024 or len(call_state.audio_buffer) >= (
-            MAXIMUM_BINARY_MESSAGE_SIZE / 2
-        ):
-            # TODO: we should send the buffer when we receive a synthesis complete event
-            # from TTS. This will ensure that the last audio chunk is always sent.
-            await self._send_audio_buffer(self.voice_websocket)
-
-    async def _send_audio_buffer(self, ws: Websocket) -> None:
-        """Send the audio buffer to the recipient if it's not empty."""
-        if call_state.audio_buffer:
-            buffer_bytes = bytes(call_state.audio_buffer)
-            await self._send_bytes_to_ws(ws, buffer_bytes)
-            call_state.audio_buffer.clear()
-
-    async def _send_bytes_to_ws(self, ws: Websocket, data: bytes) -> None:
-        """Send audio bytes to the recipient as a binary websocket message."""
-        if len(data) <= MAXIMUM_BINARY_MESSAGE_SIZE:
-            await self.voice_websocket.send(data)
-        else:
-            # split the audio into chunks
-            current_position = 0
-            while current_position < len(data):
-                end_position = min(
-                    current_position + MAXIMUM_BINARY_MESSAGE_SIZE, len(data)
-                )
-                await self.voice_websocket.send(data[current_position:end_position])
-                current_position = end_position
+        await self.voice_websocket.send(audio_bytes)
 
     async def send_marker_message(self, recipient_id: str) -> None:
         """
@@ -119,6 +89,17 @@ class GenesysInputChannel(VoiceInputChannel):
     def __init__(self, *args: Any, **kwargs: Any) -> None:
         super().__init__(*args, **kwargs)
 
+    def _ensure_channel_data_initialized(self) -> None:
+        """Initialize Genesys-specific channel data if not already present.
+
+        Genesys requires the server and client each maintain a
+        monotonically increasing message sequence number.
+        """
+        if "server_sequence_number" not in call_state.channel_data:
+            call_state.channel_data["server_sequence_number"] = 0
+        if "client_sequence_number" not in call_state.channel_data:
+            call_state.channel_data["client_sequence_number"] = 0
+
     def _get_next_sequence(self) -> int:
         """
         Get the next message sequence number
@@ -128,23 +109,26 @@ class GenesysInputChannel(VoiceInputChannel):
         Genesys requires the server and client each maintain a
         monotonically increasing message sequence number.
         """
-        cs = call_state
-        cs.server_sequence_number += 1  # type: ignore[attr-defined]
-        return cs.server_sequence_number
+        self._ensure_channel_data_initialized()
+        call_state.channel_data["server_sequence_number"] += 1
+        return call_state.channel_data["server_sequence_number"]
 
     def _get_last_client_sequence(self) -> int:
         """Get the last client(Genesys) sequence number."""
-        return call_state.client_sequence_number
+        self._ensure_channel_data_initialized()
+        return call_state.channel_data["client_sequence_number"]
 
     def _update_client_sequence(self, seq: int) -> None:
         """Update the client(Genesys) sequence number."""
-        if seq - call_state.client_sequence_number != 1:
+        self._ensure_channel_data_initialized()
+
+        if seq - call_state.channel_data["client_sequence_number"] != 1:
             logger.warning(
                 "genesys.update_client_sequence.sequence_gap",
                 received_seq=seq,
-                last_seq=call_state.client_sequence_number,
+                last_seq=call_state.channel_data["client_sequence_number"],
             )
-        call_state.client_sequence_number = seq  # type: ignore[attr-defined]
+        call_state.channel_data["client_sequence_number"] = seq
 
     def channel_bytes_to_rasa_audio_bytes(self, input_bytes: bytes) -> RasaAudioBytes:
         return RasaAudioBytes(input_bytes)
@@ -211,6 +195,7 @@ class GenesysInputChannel(VoiceInputChannel):
             voice_websocket,
             tts_engine,
             self.tts_cache,
+            min_buffer_size=MAXIMUM_BINARY_MESSAGE_SIZE // 2,
         )
 
     async def handle_open(self, ws: Websocket, message: dict) -> CallParameters: