rasa-pro 3.16.3.dev4__tar.gz → 3.16.3.dev6__tar.gz

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rasa-pro
-Version: 3.16.3.dev4
+Version: 3.16.3.dev6
 Summary: State-of-the-art open-core Conversational AI framework for Enterprises that natively leverages generative AI for effortless assistant development.
 Keywords: nlp,machine-learning,machine-learning-library,bot,bots,botkit,rasa conversational-agents,conversational-ai,chatbot,chatbot-framework,bot-framework
 Author: Rasa Technologies GmbH

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rasa-pro"
-version = "3.16.3.dev4"
+version = "3.16.3.dev6"
 description = "State-of-the-art open-core Conversational AI framework for Enterprises that natively leverages generative AI for effortless assistant development."
 authors = [ "Rasa Technologies GmbH <hi@rasa.com>",]
 maintainers = [ "Tom Bocklisch <tom@rasa.com>",]
@@ -585,12 +585,6 @@ convention = "google"
 "langfuse".msg = "Only import langfuse from rasa.builder.telemetry.langfuse_compat."
 
 
-
-
-
-
-
-
 [tool.poetry.group.dev.dependencies]
 ruff = ">=0.3.5,<0.4.0"
 docker = "~7.1.0"

rasa_pro-3.16.3.dev6/rasa/core/channels/inspector/.nvmrc

@@ -0,0 +1 @@
+24

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/core/channels/voice_stream/cfm/CFM.md

@@ -108,6 +108,15 @@ Fillers are suppressed when they'd be more annoying than helpful:
 
 Debug logging in `_should_suppress` shows which suppression rule fired and the specific values (e.g., cooldown remaining, word count, turn duration) to aid diagnosis.
 
+### Fallback Messages
+
+When the filler LLM call times out or errors, a random pre-canned message is played instead of leaving the user in silence. Both the main filler and follow-up filler have independent configurable fallback lists:
+
+- **`fallback_messages`**: For the first filler (default: "One moment.", "Sure, let me look into that.", etc.)
+- **`followup_fallback_messages`**: For follow-up fillers (default: "Bear with me.", "I appreciate your patience.", etc.)
+
+Set either list to empty (`[]`) to disable fallbacks and revert to the previous behavior (no filler on LLM failure). The default filler timeout is 2000ms — balancing speed with reliability.
+
 ---
 
 ## Follow-up Fillers
@@ -139,7 +148,13 @@ Follow-ups use `anthropic/claude-haiku-4-5-20251001` by default (configurable vi
 
 The `record_filler_on_tracker` callback records the follow-up as `BotUttered(metadata={"is_filler": True})` on the Rasa tracker. This recording happens BEFORE `play_filler_audio` to avoid `CancelledError` issues — if the task is cancelled during TTS playback (which is shielded), the tracker event is already safely recorded.
 
-**Known limitation**: Follow-up fillers appear in the `FloorContext` conversation history (used by CFM prompts for turn detection, barge-in, and subsequent fillers) but NOT in agent/command generator conversation histories built from the tracker. The follow-up `BotUttered` event lands on the tracker too late or at the wrong position for the agent's message builder to pick up.
+Follow-up fillers now reliably appear in both CFM prompts (via `FloorContext` conversation history) and agent/command generator conversation histories (via tracker events). This is achieved through two paths:
+
+**Primary path** (follow-up fires during pipeline processing — the common case): `dispatch_user_message` exposes its in-memory tracker via a shared `_active_dispatch_tracker` reference in the `_create_cfm` closure scope. When `record_filler_on_tracker` detects this reference is set, it injects the `BotUttered` event directly onto the live tracker via `tracker.update()`. The event is persisted when dispatch calls `save_tracker()` at the end of the pipeline. Ordering is naturally correct because asyncio cooperative scheduling means the follow-up appends before the response (which hasn't been generated yet): `UserUttered → BotUttered(filler) → BotUttered(followup) → BotUttered(response)`.
+
+**Fallback path** (dispatch has already finished — rare edge case): `record_filler_on_tracker` acquires the tracker lock, loads the tracker from the store, and inserts the follow-up at the correct position — right after the last `is_filler` BotUttered event and before the response. This uses direct `tracker.events.insert()` (not `tracker.update()`) for positional control. `BotUttered.apply_to` is a no-op, so direct insertion is safe.
+
+The `_active_dispatch_tracker` reference is set after filler injection (step 2 in dispatch) and cleared in the `finally` block, ensuring it's never stale.
 
 ### Cancellation Points
 
@@ -151,7 +166,7 @@ The follow-up background task is cancelled at six points:
 5. **`send_response_chunk`** — first real chunk arrives for streamed response (agent output).
 6. **After `dispatch_user_message` returns** — pipeline finished.
 
-The follow-up is intentionally NOT cancelled by empty streaming responses (agent tool calls that produce no text). This allows follow-ups to play during agent processing gaps where the user hears nothing.
+The follow-up is intentionally NOT cancelled by empty streaming responses (agent tool calls that produce no text — "ghost segments"). This allows follow-ups to play during agent processing gaps where the user hears nothing. Importantly, follow-up cancellation happens on the first **real** text chunk in `send_response_chunk` (point 5 above), NOT in `_send_response_chunk_start_locked` — because `chunk_start` fires for ghost segments too, and cancelling there would kill the follow-up during long silent processing gaps (e.g. 10+ seconds of agent tool calls) where it's needed most.
 
 `asyncio.shield` protects `play_filler_audio` from cancellation once TTS starts — the filler always plays to completion, and the real response waits on the `synthesis_lock`.
 
@@ -169,11 +184,18 @@ Agent streaming responses for tool calls can briefly set `is_bot_speaking=True`
 cfm:
   filler:
     model: openai/gpt-5.4-mini
+    model_timeout_ms: 2000                    # default: 2000
+    fallback_messages:                        # played on LLM timeout/error
+      - "One moment."
+      - "Sure, let me look into that."
     followup_enabled: true                    # default: true
     followup_delay_seconds: 2.0               # seconds after first filler ends
     followup_model: anthropic/claude-haiku-4-5-20251001
     followup_model_timeout_ms: 2500
     followup_prompt_template: prompts/custom_followup.jinja2  # optional
+    followup_fallback_messages:               # played on follow-up LLM timeout/error
+      - "Bear with me."
+      - "I appreciate your patience."
 ```
 
 ---

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/core/channels/voice_stream/cfm/config.py

@@ -3,7 +3,7 @@
 from __future__ import annotations
 
 from dataclasses import dataclass, field
-from typing import Optional
+from typing import List, Optional
 
 # ---------------------------------------------------------------------------
 # Turn Detection
@@ -107,9 +107,21 @@ class FillerConfig:
     model_api_base: Optional[str] = None
     model_api_key: Optional[str] = None
     # Maximum time (ms) to wait for the LLM response.
-    model_timeout_ms: int = 1500
+    model_timeout_ms: int = 2000
     # Temperature — slightly above 0 for variety in fillers.
     model_temperature: float = 0.3
+    # Pre-canned fallback messages used when the LLM filler call times out
+    # or errors.  A random message is picked from this list.  If empty,
+    # no filler is played on LLM failure (previous behavior).
+    fallback_messages: List[str] = field(
+        default_factory=lambda: [
+            "One moment.",
+            "Sure, let me look into that.",
+            "Let me check on that for you.",
+            "Bear with me.",
+            "Sure, one moment.",
+        ]
+    )
 
     # -- Prompt template ----
     # Path to a custom Jinja2 template file for the system prompt.
@@ -148,6 +160,16 @@ class FillerConfig:
     # Path to a custom Jinja2 template file for the follow-up filler prompt.
     # If None, uses the built-in default template.
     followup_prompt_template: Optional[str] = None
+    # Pre-canned fallback messages for follow-up fillers when the LLM call
+    # times out or errors.  A random message is picked from this list.
+    # If empty, no follow-up is played on LLM failure.
+    followup_fallback_messages: List[str] = field(default_factory=lambda: [
+        "Bear with me.",
+        "I appreciate your patience.",
+        "Still working on that.",
+        "Almost there.",
+        "Just a moment longer.",
+    ])
 
 
 # ---------------------------------------------------------------------------

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/core/channels/voice_stream/cfm/floor_manager.py

@@ -479,7 +479,7 @@ class ConversationFloorManager:
             # Start both LLM calls concurrently.
             try:
                 turn_eval, filler_text = await asyncio.gather(
-                    _timed_turn_detection(), # type: ignore[no-untyped-call]
+                    _timed_turn_detection(),  # type: ignore[no-untyped-call]
                     _timed_filler_generation(),  # type: ignore[no-untyped-call]
                 )
             except asyncio.CancelledError:

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/core/channels/voice_stream/filler/filler_generator.py

@@ -13,6 +13,7 @@ from __future__ import annotations
 import asyncio
 import importlib.resources
 import logging
+import random
 import time as _time
 from typing import List, Optional, Tuple
 
@@ -254,7 +255,7 @@ class FillerGenerator:
                 type(e).__name__,
                 str(e)[:200] or "(no message)",
             )
-            return None
+            return self._pick_followup_fallback("followup_" + type(e).__name__)
 
     # -----------------------------------------------------------------
     # Suppression rules
@@ -308,6 +309,30 @@ class FillerGenerator:
 
         return False
 
+    def _pick_fallback(self, reason: str) -> Optional[str]:
+        """Pick a random pre-canned fallback message, or None if none configured."""
+        if not self.config.fallback_messages:
+            return None
+        fallback = random.choice(self.config.fallback_messages)
+        logger.info(
+            "filler_generator.using_fallback: reason=%s fallback=%s",
+            reason,
+            fallback,
+        )
+        return fallback
+
+    def _pick_followup_fallback(self, reason: str) -> Optional[str]:
+        """Pick a random pre-canned follow-up fallback, or None if none configured."""
+        if not self.config.followup_fallback_messages:
+            return None
+        fallback = random.choice(self.config.followup_fallback_messages)
+        logger.info(
+            "filler_generator.using_followup_fallback: reason=%s fallback=%s",
+            reason,
+            fallback,
+        )
+        return fallback
+
     # -----------------------------------------------------------------
     # LLM generation
     # -----------------------------------------------------------------
@@ -404,7 +429,7 @@ class FillerGenerator:
                 "filler_generator.llm_call_timeout: timeout_ms=%d",
                 self.config.model_timeout_ms,
             )
-            return None
+            return self._pick_fallback("timeout")
 
         except Exception as e:
             latency_ms = (_time.perf_counter() - t0) * 1000
@@ -414,4 +439,4 @@ class FillerGenerator:
                 str(e)[:200],
                 latency_ms,
             )
-            return None
+            return self._pick_fallback("error")

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/core/channels/voice_stream/voice_channel.py

@@ -743,10 +743,11 @@ class VoiceOutputChannel(OutputChannel):
         self, recipient_id: Text, **kwargs: Any
     ) -> None:
         """Inner send_response_chunk_start, called with synthesis_lock held."""
-        # Cancel follow-up filler early — before the pacing delay — so
-        # that it doesn't play during the gap between filler and response.
-        if self._cfm and not self._is_filler:
-            self._cfm._cancel_followup_filler()
+        # Follow-up filler cancellation is deferred to send_response_chunk
+        # (first real text chunk) rather than here, because chunk_start
+        # also fires for ghost segments (empty agent tool call iterations).
+        # Cancelling here would kill the follow-up during long silent
+        # processing gaps where it's actually needed.
         await self._apply_min_delay_between_messages(recipient_id)
 
         # Set TTS start time right after pacing delay, before TTS setup,
@@ -1419,6 +1420,12 @@ class VoiceInputChannel(InputChannel):
             ConversationFloorManager,
         )
 
+        # Shared reference to the in-memory tracker held by dispatch_user_message.
+        # When set, record_filler_on_tracker injects directly onto this tracker
+        # (which dispatch will save), avoiding the stale-snapshot overwrite race.
+        _active_dispatch_tracker = None
+        _active_dispatch_domain = None
+
         @observe(
             name="VoiceInputChannel._create_cfm.dispatch_user_message",
             capture_input=False,
@@ -1492,6 +1499,12 @@ class VoiceInputChannel(InputChannel):
                             )
                             tracker.update(filler_event, processor.domain)
 
+                            # Expose tracker so follow-up fillers can inject
+                            # directly (avoids stale-snapshot overwrite race).
+                            nonlocal _active_dispatch_tracker, _active_dispatch_domain
+                            _active_dispatch_tracker = tracker
+                            _active_dispatch_domain = processor.domain
+
                             # Step 3: run the pipeline (produces the real response).
                             if (
                                 processor.model_metadata.training_type
@@ -1511,6 +1524,8 @@ class VoiceInputChannel(InputChannel):
                             await out_ch.notify_message_processing_completed()
                             processor.trigger_anonymization(tracker)
                         finally:
+                            _active_dispatch_tracker = None
+                            _active_dispatch_domain = None
                             processor.unregister_cancellation_token(sender_id)
                 else:
                     await on_new_message(message)
@@ -1534,23 +1549,75 @@ class VoiceInputChannel(InputChannel):
 
             Used for follow-up fillers that aren't part of the
             dispatch_user_message flow.
+
+            Two paths:
+              1. **Primary** — dispatch_user_message is still running and
+                 holds the tracker lock.  We inject directly onto its
+                 in-memory tracker; dispatch's final save_tracker() will
+                 persist the event.  Natural append order is correct
+                 because the pipeline response hasn't been generated yet.
+              2. **Fallback** — dispatch has already finished.  We acquire
+                 the lock ourselves, load the tracker, and insert the
+                 follow-up at the correct position (after the last filler,
+                 before the response) so future agent turns see the right
+                 conversation history.
             """
+            nonlocal _active_dispatch_tracker, _active_dispatch_domain
+            from rasa.shared.core.events import BotUttered
+
+            filler_event = BotUttered(
+                text=filler_text,
+                metadata={"is_filler": True},
+            )
+
+            # --- Primary path: inject onto the live dispatch tracker ---
+            if _active_dispatch_tracker is not None:
+                _active_dispatch_tracker.update(
+                    filler_event, _active_dispatch_domain
+                )
+                logger.debug(
+                    "voice_channel.record_filler_on_tracker: "
+                    "injected on active dispatch tracker '%s'",
+                    filler_text[:40],
+                )
+                return
+
+            # --- Fallback: dispatch finished, load with lock ---
             if not self.agent or not self.agent.processor:
                 logger.debug("voice_channel.record_filler_on_tracker: no processor")
                 return
             sender_id = self.get_sender_id(call_parameters)
-            from rasa.shared.core.events import BotUttered
+            processor = self.agent.processor
+
+            async with self.agent.lock_store.lock(sender_id):
+                tracker = await processor.get_tracker(sender_id)
+                if not tracker:
+                    return
+
+                # Find insertion point: right after the last is_filler
+                # BotUttered, so the order is filler → followup → response.
+                insert_idx = len(tracker.events)
+                for i in range(len(tracker.events) - 1, -1, -1):
+                    evt = tracker.events[i]
+                    if isinstance(evt, BotUttered):
+                        if evt.metadata and evt.metadata.get("is_filler"):
+                            insert_idx = i + 1
+                            break
+                        else:
+                            # Non-filler BotUttered (the response) — insert
+                            # before it.
+                            insert_idx = i
+                            break
 
-            tracker = await self.agent.processor.get_tracker(sender_id)
-            if tracker:
-                filler_event = BotUttered(
-                    text=filler_text,
-                    metadata={"is_filler": True},
-                )
-                tracker.update(filler_event, self.agent.processor.domain)
-                await self.agent.processor.save_tracker(tracker)
+                # Direct insert (not tracker.update) to avoid append
+                # semantics and keep the event at the correct position.
+                filler_event.timestamp = time.time()
+                tracker.events.insert(insert_idx, filler_event)
+                await processor.save_tracker(tracker)
                 logger.debug(
-                    "voice_channel.record_filler_on_tracker: recorded '%s'",
+                    "voice_channel.record_filler_on_tracker: "
+                    "inserted at idx=%d (fallback) '%s'",
+                    insert_idx,
                     filler_text[:40],
                 )
 

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/dialogue_understanding/patterns/default_flows_for_patterns.yml

@@ -247,7 +247,7 @@ flows:
       - id: action_clarify_flows
         action: action_clarify_flows
       - action: utter_clarification_options_rasa
-  
+
   pattern_code_change:
     description: Conversation repair flow for cleaning the stack after an assistant update
     name: pattern code change

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/shared/core/flows/flows_yaml_schema.json

@@ -150,6 +150,7 @@
                             "metadata": {},
                             "description": {"type": "string"},
                             "ask_before_filling": {"type": "boolean"},
+                            "silent_collect": {"type": "boolean"},
                             "reset_after_flow_ends": {"type": "boolean"},
                             "utter": {"type": "string"},
                             "force_slot_filling": {"type": "boolean"},

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/shared/core/flows/steps/collect.py

@@ -15,6 +15,7 @@ from rasa.shared.exceptions import RasaException
 DEFAULT_ASK_BEFORE_FILLING = False
 DEFAULT_RESET_AFTER_FLOW_ENDS = True
 DEFAULT_FORCE_SLOT_FILLING = False
+DEFAULT_SILENT_COLLECT = False
 
 logger = structlog.get_logger(__name__)
 
@@ -175,6 +176,14 @@ class CollectInformationFlowStep(FlowStep):
     """The silence timeout for the collect information step."""
     dtmf: Optional[DTMFConfig] = None
     """DTMF configuration for the collect information step."""
+    silent_collect: bool = DEFAULT_SILENT_COLLECT
+    """If True, this step may be unreachable; it only registers the slot for CALM.
+
+    Silent collect steps are still listed in ``get_collect_steps()`` so the command
+    generator can fill the slot (typically with ``ask_before_filling: false``) without
+    the flow graph requiring an incoming edge to this step. The flow executor never
+    navigates to an unreachable step.
+    """
 
     @classmethod
     def from_json(
@@ -214,6 +223,7 @@ class CollectInformationFlowStep(FlowStep):
             force_slot_filling=data.get("force_slot_filling", False),
             silence_timeout=silence_timeout,
             dtmf=dtmf,
+            silent_collect=data.get("silent_collect", DEFAULT_SILENT_COLLECT),
             **base.__dict__,
         )
 
@@ -260,6 +270,8 @@ class CollectInformationFlowStep(FlowStep):
             data.update(self.silence_timeout.to_json())
         if self.dtmf:
             data["dtmf"] = self.dtmf.to_json()
+        if self.silent_collect:
+            data["silent_collect"] = True
 
         return super().as_json(step_properties=data)
 
@@ -285,6 +297,7 @@ class CollectInformationFlowStep(FlowStep):
                 and self.force_slot_filling == other.force_slot_filling
                 and self.silence_timeout == other.silence_timeout
                 and self.dtmf == other.dtmf
+                and self.silent_collect == other.silent_collect
                 and super().__eq__(other)
             )
         return False
@@ -302,6 +315,7 @@ class CollectInformationFlowStep(FlowStep):
                 self.force_slot_filling,
                 self.silence_timeout,
                 self.dtmf,
+                self.silent_collect,
                 self.idx,
                 self.description,
                 self.flow_id,

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/shared/core/flows/validation.py

@@ -64,7 +64,28 @@ class UnreachableFlowStepException(RasaException):
             f"Step '{self.step_id}' in flow '{self.flow_id}' can not be reached "
             f"from the start step. Please make sure that all steps can be reached "
             f"from the start step, e.g. by "
-            f"checking that another step points to this step."
+            f"checking that another step points to this step. "
+            f"Collect steps that only register a slot for CALM (no incoming edge) "
+            f"may use 'silent_collect: true'."
+        )
+
+
+class SilentCollectAsStartStepException(RasaException):
+    """Raised when the first step of a flow is a silent collect step."""
+
+    def __init__(self, flow_id: str, step_id: str) -> None:
+        """Initializes the exception."""
+        self.flow_id = flow_id
+        self.step_id = step_id
+
+    def __str__(self) -> str:
+        """Return a string representation of the exception."""
+        return (
+            f"Step '{self.step_id}' in flow '{self.flow_id}' uses "
+            f"'silent_collect: true' "
+            f"but it is the first step in the flow. Silent collect steps must appear "
+            f"after at least one reachable step so they only register slots for CALM "
+            f"and are not used as the flow entry point."
         )
 
 
@@ -567,6 +588,7 @@ def validate_flow(flow: Flow) -> None:
     validate_all_steps_next_property(flow)
     validate_all_next_ids_are_available_steps(flow)
     validate_all_steps_can_be_reached(flow)
+    validate_silent_collect_not_start_step(flow)
     validate_all_branches_have_an_else(flow)
     validate_not_using_builtin_ids(flow)
     validate_slot_names_to_be_collected(flow)
@@ -658,9 +680,18 @@ def validate_all_steps_can_be_reached(flow: Flow) -> None:
 
     for step in flow.steps:
         if step.id not in reached_steps:
+            if isinstance(step, CollectInformationFlowStep) and step.silent_collect:
+                continue
             raise UnreachableFlowStepException(step.id, flow.id)
 
 
+def validate_silent_collect_not_start_step(flow: Flow) -> None:
+    """Validate that a silent collect step is not the flow entry step."""
+    first = flow.first_step_in_flow()
+    if isinstance(first, CollectInformationFlowStep) and first.silent_collect:
+        raise SilentCollectAsStartStepException(flow.id, first.id)
+
+
 def validate_nlu_trigger(flows: List[Flow]) -> None:
     """Validates that an intent can just trigger one flow."""
     nlu_trigger_to_flows = defaultdict(list)

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/shared/core/flows/yaml_flows_io.py

@@ -16,6 +16,7 @@ from rasa.shared.core.flows.steps.collect import (
     DEFAULT_ASK_BEFORE_FILLING,
     DEFAULT_FORCE_SLOT_FILLING,
     DEFAULT_RESET_AFTER_FLOW_ENDS,
+    DEFAULT_SILENT_COLLECT,
 )
 from rasa.shared.exceptions import RasaException, YamlException
 from rasa.shared.utils.yaml import (
@@ -507,6 +508,8 @@ def get_flow_as_json(flow: Flow, should_clean_json: bool = False) -> Dict[str, A
             step_data.pop("reset_after_flow_ends", None)
         if step_data.get("force_slot_filling") is DEFAULT_FORCE_SLOT_FILLING:
             step_data.pop("force_slot_filling", None)
+        if step_data.get("silent_collect") is DEFAULT_SILENT_COLLECT:
+            step_data.pop("silent_collect", None)
         if not step_data.get("rejections"):
             step_data.pop("rejections", None)
 

{rasa_pro-3.16.3.dev4 → rasa_pro-3.16.3.dev6}/rasa/version.py

@@ -1,3 +1,3 @@
 # this file will automatically be changed,
 # do not add anything but the version number here!
-__version__ = "3.16.3.dev4"
+__version__ = "3.16.3.dev6"

rasa_pro-3.16.3.dev4/rasa/core/channels/inspector/.nvmrc

@@ -1 +0,0 @@
-20.19.2