voice-agent-examples / src /nvidia_pipecat /processors /nvidia_context_aggregator.py
fciannella's picture
Working with service run on 7860
53ea588
raw
history blame
21.7 kB
# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: BSD 2-Clause License
"""NVIDIA Context Aggregator.
This module provides specialized frame processors and context aggregators for
handling NVIDIA's Speculative Speech Processing feature in conversational AI systems.
It manages the processing and aggregation of interim and final transcripts,
enabling real-time response generation while maintaining conversation coherence.
The processors handle:
- Interim transcript processing for early response generation
- Context management for speculative responses
- TTS response caching and timing control for natural turn-taking
- Bidirectional conversation state management
Also see:
pipecat.processors.aggregators.llm_response
pipecat.processors.aggregators.openai_llm_context
Classes:
NvidiaAssistantContextAggregator: Handles assistant-specific context aggregation.
NvidiaUserContextAggregator: Manages user context with interim/final transcripts.
NvidiaTTSResponseCacher: Controls TTS response timing.
NvidiaContextAggregatorPair: Coordinates paired aggregators.
Functions:
create_nvidia_context_aggregator: Factory for creating aggregator pairs.
"""
from dataclasses import dataclass
from loguru import logger
from pipecat.frames.frames import (
Frame,
LLMFullResponseEndFrame,
LLMFullResponseStartFrame,
StartInterruptionFrame,
TranscriptionFrame,
TTSAudioRawFrame,
TTSStartedFrame,
TTSStoppedFrame,
TTSTextFrame,
UserStartedSpeakingFrame,
UserStoppedSpeakingFrame,
)
from pipecat.processors.aggregators.llm_response import (
LLMAssistantAggregatorParams,
LLMAssistantContextAggregator,
LLMUserAggregatorParams,
LLMUserContextAggregator,
)
from pipecat.processors.aggregators.openai_llm_context import OpenAILLMContext, OpenAILLMContextFrame
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from nvidia_pipecat.frames.action import StartedPresenceUserActionFrame
from nvidia_pipecat.frames.riva import RivaInterimTranscriptionFrame
class NvidiaAssistantContextAggregator(LLMAssistantContextAggregator):
"""Extends LLMAssistantContextAggregator for NVIDIA-specific requirements.
Specializes the base aggregator for handling speculative speech processing,
managing assistant responses and context updates based on interim/final transcripts.
Args:
context (OpenAILLMContext): The context object to use.
expect_stripped_words (bool): Whether to expect preprocessed words. Defaults to True.
**kwargs: Additional arguments passed to parent class.
Input Frames:
LLMFullResponseStartFrame: Marks response start
LLMFullResponseEndFrame: Marks response end
TextFrame: Contains response text
StartInterruptionFrame: Signals interruption
Output Frames:
OpenAILLMContextFrame: Updated context with responses
"""
async def push_aggregation(self):
"""Updates the context with current aggregation.
For speculative processing, may update existing messages rather than append
to maintain context coherence with interim transcripts.
- If the last message in context has the same role, it updates that message
- Otherwise, appends a new message with the current aggregation
- After pushing, resets the aggregation state
Returns:
None
Typical usage example:
>>> context = OpenAILLMContext()
>>> aggregator = NvidiaAssistantContextAggregator(context)
>>> # Update existing response
>>> context.add_message({"role": "assistant", "content": "initial response"})
>>> aggregator._aggregation = "updated response"
>>> await aggregator.push_aggregation()
"""
if len(self._aggregation) > 0:
context_messages = self.context.get_messages()
# Update existing message if same role, otherwise append new one
if len(context_messages) > 0 and context_messages[-1]["role"] == self._role:
context_messages[-1]["content"] = self._aggregation
self.context.set_messages(context_messages)
else:
self.context.add_message({"role": self._role, "content": self._aggregation})
self._aggregation = ""
frame = OpenAILLMContextFrame(self.context)
await self.push_frame(frame)
# Reset our accumulator state.
self.reset()
class NvidiaUserContextAggregator(LLMUserContextAggregator):
"""Extends LLMUserContextAggregator for user-specific context handling.
Handles speculative speech processing with interim and final transcriptions.
Key features for speculative processing:
- Processes stable interim transcripts for early response generation
- Manages transition from interim to final transcripts
- Deduplicates repeated transcripts to prevent context pollution
- Maintains conversation history with configurable turn limits
- Tracks user speaking state to coordinate with assistant responses
Input Frames:
TranscriptionFrame: Final transcription
RivaInterimTranscriptionFrame: Interim transcription
UserStartedSpeakingFrame: User began speaking
UserStoppedSpeakingFrame: User stopped speaking
StartInterruptionFrame: Conversation interruption
Output Frames:
OpenAILLMContextFrame: Updated context with transcripts
"""
def __init__(
self,
send_interims: bool = True,
chat_history_limit: int = 20,
**kwargs,
):
"""Initialize the NvidiaUserContextAggregator.
Args:
send_interims (bool, optional): Whether to send interim transcription frames. Defaults to True.
chat_history_limit (int): Limits the number of turns in chat history,
**kwargs: Additional keyword arguments passed to parent LLMUserContextAggregator.
"""
super().__init__(**kwargs)
self.send_interims = send_interims
self.chat_history_limit = chat_history_limit
self.last_transcript = None
self._user_speaking = False
self.seen_final = True
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process a frame for speculative speech handling.
- Processes stable interim transcripts when user is speaking
- Manages transition between interim and final transcripts
- Handles user speaking state changes
- Deduplicates repeated transcripts
Args:
frame: Frame to process (TranscriptionFrame, RivaInterimTranscriptionFrame,
or user state frames)
direction: Direction of frame flow in the pipeline
Typical usage example:
>>> aggregator = NvidiaUserContextAggregator(
... send_interims=True, # Enable interim transcript processing
... chat_history_limit=20 # Keep last 20 conversation turns
... )
>>> # Process final transcript
>>> frame = TranscriptionFrame(text="Hello")
>>> await aggregator.process_frame(frame, FrameDirection.DOWNSTREAM)
>>>
>>> # Process interim transcript
>>> frame = RivaInterimTranscriptionFrame(text="Hello", stability=1.0)
>>> await aggregator.process_frame(frame, FrameDirection.DOWNSTREAM)
"""
if isinstance(frame, TranscriptionFrame):
logger.debug(f"Recieved final transcript at NvidiaUserContextAggregator {frame.text}")
# Only process if this is a new transcript
if self.last_transcript is None or (self.last_transcript.rstrip() != frame.text.rstrip()):
logger.debug(f"Sent final transcript downstream to LLM from NvidiaUserContextAggregator {frame.text}")
self._aggregation = frame.text
await self.push_aggregation()
self.last_transcript = None
self.seen_final = True
elif isinstance(frame, RivaInterimTranscriptionFrame):
# Process stable interim transcriptions during active speech and before first final result
if self.send_interims and (self._user_speaking or not self.seen_final) and frame.stability == 1.0:
logger.debug(f"Sent interim transcript downstream to LLM from NvidiaUserContextAggregator {frame.text}")
self._aggregation = frame.text
await self.push_aggregation()
self.last_transcript = frame.text
elif isinstance(frame, StartInterruptionFrame):
self._user_speaking = False
await self._start_interruption()
await self.stop_all_metrics()
await self.push_frame(frame, direction)
else:
if isinstance(frame, UserStartedSpeakingFrame):
self._user_speaking = True
self.seen_final = False
elif isinstance(frame, UserStoppedSpeakingFrame):
self._user_speaking = False
await super().process_frame(frame, direction)
async def get_truncated_context(self) -> OpenAILLMContext:
"""Returns a truncated context limited to specified chat history size.
- Counts conversation turns based on user-assistant exchanges
- Preserves system and function messages regardless of limit
- Processes messages in reverse order to maintain recent history
Returns:
OpenAILLMContext: New context object containing truncated conversation
history, preserving system/function messages and most recent turns.
Typical usage example:
>>> aggregator = NvidiaUserContextAggregator(chat_history_limit=2)
>>> # Context with 3 turns
>>> context = OpenAILLMContext()
>>> # Turn 1
>>> context.add_message({"role": "user", "content": "Turn 1 user"})
>>> context.add_message({"role": "assistant", "content": "Turn 1 assistant"})
>>> # Turn 2
>>> context.add_message({"role": "user", "content": "Turn 2 user"})
>>> context.add_message({"role": "assistant", "content": "Turn 2 assistant"})
>>> # Turn 3
>>> context.add_message({"role": "user", "content": "Turn 3 user"})
>>> context.add_message({"role": "assistant", "content": "Turn 3 assistant"})
>>> # Get truncated context - will only contain most recent 2 turns
>>> truncated = await aggregator.get_truncated_context()
>>> print(truncated.get_messages()) # Shows turns 2 and 3 only
"""
truncated_context = self.context
if len(self.context.get_messages()) > 0:
truncated_context = OpenAILLMContext()
truncated_context_messages = []
current_size = 0
for context_message in reversed(self.context.get_messages()):
if (
context_message["role"] == "user"
or context_message["role"] == "assistant"
or context_message["role"] == "developer"
or context_message["role"] == "function"
or context_message["role"] == "tool"
):
if current_size == self.chat_history_limit:
continue
if context_message["role"] == "user":
current_size = current_size + 1
truncated_context_messages.append(context_message)
truncated_context.set_messages(reversed(truncated_context_messages))
return truncated_context
async def push_aggregation(self):
"""Pushes aggregation to context and manages conversation flow.
Updates or appends current aggregation to conversation context while
maintaining turn-taking structure. For speculative responses, may update
existing message rather than append new one.
- If the last message in context has the same role, it updates that message
- Otherwise, appends a new message with the current aggregation
- After pushing, resets the aggregation state
Output Frames:
OpenAILLMContextFrame: downstream after processing.
Typical usage example:
>>> context = OpenAILLMContext()
>>> aggregator = NvidiaUserContextAggregator(context)
>>> # Update existing response
>>> context.add_message({"role": "user", "content": "initial query"})
>>> aggregator._aggregation = "updated query"
>>> await aggregator.push_aggregation()
"""
if len(self._aggregation) > 0:
context_messages = self.context.get_messages()
# Update existing message if same role, otherwise append new one
if len(context_messages) > 0 and context_messages[-1]["role"] == self._role:
context_messages[-1]["content"] = self._aggregation
self.context.set_messages(context_messages)
else:
self.context.add_message({"role": self._role, "content": self._aggregation})
self._aggregation = ""
# Get truncated context and send downstream
truncated_context = await self.get_truncated_context()
frame = OpenAILLMContextFrame(truncated_context)
# Send the interruption before the context frame
await self.push_frame(StartInterruptionFrame())
logger.debug(
f"Sending context downstream to LLM from NvidiaUserContextAggregator {frame.context.get_messages()}"
)
await self.push_frame(frame)
# Reset our accumulator state
self.reset()
class NvidiaTTSResponseCacher(FrameProcessor):
"""Caches TTS responses and controls release timing for speculative speech.
Manages text-to-speech response timing by caching responses and controlling
their release based on user speaking state. Maintains natural turn-taking
and prevents response overlap during speculative processing.
Input frames handled:
- LLMFullResponseStartFrame: Marks response start
- LLMFullResponseEndFrame: Marks response end
- TTSAudioRawFrame: TTS audio data
- TTSStartedFrame: TTS start marker
- TTSStoppedFrame: TTS stop marker
- TTSTextFrame: TTS text data
- UserStartedSpeakingFrame: Triggers caching
- UserStoppedSpeakingFrame: Triggers release
- StartInterruptionFrame: Clears cache
"""
def __init__(self) -> None:
"""Initialize the NvidiaTTSResponseCacher."""
super().__init__()
self._cache: list[Frame] = []
self.user_stopped_speaking: bool = True
async def process_frame(self, frame: Frame, direction: FrameDirection) -> None:
"""Processes frame for TTS response caching and timing control.
- Caches TTS responses while user is speaking
- Releases cached responses when user stops speaking
- Clears cache on interruptions
- Maintains conversation flow by coordinating response timing
Also see:
- NvidiaUserContextAggregator : Handles user context and speech state
- NvidiaAssistantContextAggregator : Manages assistant responses
Args:
frame: Frame to process
direction: Direction of frame flow in pipeline
Typical usage example:
>>> cacher = NvidiaTTSResponseCacher()
>>> # User starts speaking
>>> await cacher.process_frame(UserStartedSpeakingFrame(), FrameDirection.DOWNSTREAM)
>>> # TTS response arrives - will be cached
>>> await cacher.process_frame(TTSAudioRawFrame(audio_data), FrameDirection.DOWNSTREAM)
>>> # User stops speaking - cached responses will be released
>>> await cacher.process_frame(UserStoppedSpeakingFrame(), FrameDirection.DOWNSTREAM)
"""
await super().process_frame(frame, direction)
# Handle response start - cache if user is speaking
if isinstance(frame, LLMFullResponseStartFrame):
if self.user_stopped_speaking:
await self.push_frame(frame, direction)
else:
self._cache.clear() # Clear existing cache before new response
self._cache.append(frame)
# Handle TTS frames - cache or forward based on user speaking state
elif isinstance(frame, (TTSAudioRawFrame | TTSStartedFrame | TTSStoppedFrame | TTSTextFrame)):
if self.user_stopped_speaking:
await self.push_frame(frame, direction)
else:
self._cache.append(frame)
# Handle response end - mark user can speak after forwarding
elif isinstance(frame, LLMFullResponseEndFrame):
if self.user_stopped_speaking:
await self.push_frame(frame, direction)
self.user_stopped_speaking = False # Allow user to speak after response ends
self._cache.append(frame)
# Handle interruptions - clear cache and reset state
elif isinstance(frame, StartInterruptionFrame | StartedPresenceUserActionFrame):
# TODO: This only works if we have a single user in the system.
# it also does not work if other "events" should trigger the cache release
# (e.g. new frames by new processors).
self._cache.clear()
# self.user_stopped_speaking = True
await self.push_frame(frame, direction)
# Handle user stop speaking - release cached responses
elif isinstance(frame, UserStoppedSpeakingFrame):
self.user_stopped_speaking = True
if self._cache:
for cached_frame in self._cache:
await self.push_frame(cached_frame)
self._cache.clear()
await self.push_frame(frame, direction)
# Handle user start speaking - update state
elif isinstance(frame, UserStartedSpeakingFrame):
self.user_stopped_speaking = False
await self.push_frame(frame, direction)
# Forward all other frames unchanged
else:
await self.push_frame(frame, direction)
@dataclass
class NvidiaContextAggregatorPair:
"""A pair of context aggregators for managing bidirectional conversation.
Attributes:
_user: NvidiaUserContextAggregator for user-side context
_assistant: NvidiaAssistantContextAggregator for assistant-side context
"""
_user: "NvidiaUserContextAggregator"
_assistant: "NvidiaAssistantContextAggregator"
def user(self) -> "NvidiaUserContextAggregator":
"""Get the user context aggregator."""
return self._user
def assistant(self) -> "NvidiaAssistantContextAggregator":
"""Get the assistant context aggregator."""
return self._assistant
def create_nvidia_context_aggregator(
context: OpenAILLMContext,
assistant_expect_stripped_words: bool = True,
send_interims: bool = False,
chat_history_limit: int = 20,
) -> NvidiaContextAggregatorPair:
"""Creates a pair of context aggregators for speculative speech processing.
- Creates synchronized user and assistant aggregators sharing context
- User aggregator handles interim/final transcripts
- Assistant aggregator manages response generation
- Both work together to maintain conversation coherence
Also see:
- NvidiaUserContextAggregator : Handles user context
- NvidiaAssistantContextAggregator : Handles assistant context
Args:
context: Base context object to initialize aggregators
assistant_expect_stripped_words: Whether assistant expects preprocessed words
send_interims: Whether to process interim transcriptions
chat_history_limit: Maximum number of conversation turns to maintain
Returns:
NvidiaContextAggregatorPair: A paired set of user and assistant context aggregators configured for
speculative speech processing.
Typical usage example:
>>> context = OpenAILLMContext()
>>> # Create aggregators with default settings
>>> aggregators = create_nvidia_context_aggregator(context)
>>>
>>> # Create aggregators with custom settings
>>> aggregators = create_nvidia_context_aggregator(
... context,
... send_interims=True, # Enable interim transcript processing
... chat_history_limit=10, # Keep shorter history
... assistant_expect_stripped_words=False # Raw word processing
... )
>>>
>>> # Access individual aggregators
>>> user_aggregator = aggregators.user()
>>> assistant_aggregator = aggregators.assistant()
"""
# Create user aggregator with specified settings
user_params = LLMUserAggregatorParams(aggregation_timeout=0.01)
user = NvidiaUserContextAggregator(
send_interims=send_interims, context=context, params=user_params, chat_history_limit=chat_history_limit
)
# Create assistant aggregator sharing context with user
assistant_params = LLMAssistantAggregatorParams(expect_stripped_words=assistant_expect_stripped_words)
assistant = NvidiaAssistantContextAggregator(context=user.context, params=assistant_params)
return NvidiaContextAggregatorPair(_user=user, _assistant=assistant)