Save changes
This commit is contained in:
James Hush
@@ -12,12 +12,17 @@ from loguru import logger
|
||||
from openai.types.chat import ChatCompletionMessageParam
|
||||
|
||||
from pipecat.audio.vad.silero import SileroVADAnalyzer
|
||||
from pipecat.frames.frames import Frame, LLMMessagesFrame
|
||||
from pipecat.frames.frames import Frame, LLMMessagesFrame, LLMMessagesUpdateFrame, LLMTextFrame
|
||||
from pipecat.observers.loggers.debug_log_observer import DebugLogObserver, FrameEndpoint
|
||||
from pipecat.observers.loggers.llm_log_observer import LLMLogObserver
|
||||
from pipecat.pipeline.parallel_pipeline import ParallelPipeline
|
||||
from pipecat.pipeline.pipeline import Pipeline
|
||||
from pipecat.pipeline.runner import PipelineRunner
|
||||
from pipecat.pipeline.task import PipelineParams, PipelineTask
|
||||
from pipecat.processors.aggregators.openai_llm_context import OpenAILLMContext
|
||||
from pipecat.processors.aggregators.openai_llm_context import (
|
||||
OpenAILLMContext,
|
||||
OpenAILLMContextFrame,
|
||||
)
|
||||
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
|
||||
from pipecat.runner.types import RunnerArguments
|
||||
from pipecat.runner.utils import create_transport
|
||||
@@ -53,83 +58,45 @@ transport_params = {
|
||||
|
||||
|
||||
class LLMRaceProcessor(FrameProcessor):
|
||||
"""Processor that sends frames to two LLMs in parallel and uses the first response."""
|
||||
"""Manages racing between two LLMs - only allows frames from the first LLM to respond."""
|
||||
|
||||
def __init__(self, llm1: OpenAILLMService, llm2: OpenAILLMService):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self._llm1 = llm1
|
||||
self._llm2 = llm2
|
||||
self._race_counter = 0
|
||||
self._active_races = {} # race_id -> winner_name
|
||||
self._current_llm_name = None
|
||||
|
||||
async def process_frame(self, frame, direction):
|
||||
await super().process_frame(frame, direction)
|
||||
|
||||
# Check if this is a frame we should race through both LLMs
|
||||
if isinstance(frame, LLMMessagesFrame):
|
||||
race_id = self._race_counter
|
||||
self._race_counter += 1
|
||||
|
||||
logger.info(f"[LLM_RACE {race_id}] Starting parallel processing")
|
||||
|
||||
# Create a result collector for this race
|
||||
race_result = asyncio.Event()
|
||||
winning_frames = []
|
||||
|
||||
async def llm_runner(llm: OpenAILLMService, name: str):
|
||||
"""Run LLM and collect results."""
|
||||
try:
|
||||
# Create a frame collector
|
||||
class FrameCollector(FrameProcessor):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.collected_frames = []
|
||||
|
||||
async def process_frame(self, frame, direction):
|
||||
self.collected_frames.append(frame)
|
||||
|
||||
collector = FrameCollector()
|
||||
|
||||
# Temporarily link LLM to collector
|
||||
llm.link(collector)
|
||||
|
||||
# Process the frame
|
||||
await llm.process_frame(frame, FrameDirection.DOWNSTREAM)
|
||||
|
||||
# Check if we won the race
|
||||
if race_id not in self._active_races:
|
||||
self._active_races[race_id] = name
|
||||
winning_frames.extend(collector.collected_frames)
|
||||
race_result.set()
|
||||
logger.info(
|
||||
f"[LLM_RACE {race_id}] {name} WON with {len(collector.collected_frames)} frames!"
|
||||
)
|
||||
else:
|
||||
logger.info(f"[LLM_RACE {race_id}] {name} lost")
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"[LLM_RACE {race_id}] Error in {name}: {e}")
|
||||
|
||||
# Start both LLMs racing
|
||||
task1 = asyncio.create_task(llm_runner(self._llm1, "LLM1"))
|
||||
task2 = asyncio.create_task(llm_runner(self._llm2, "LLM2"))
|
||||
|
||||
# Wait for the first one to complete
|
||||
await race_result.wait()
|
||||
|
||||
# Cancel the slower task
|
||||
task1.cancel()
|
||||
task2.cancel()
|
||||
|
||||
# Push the winning frames
|
||||
for winning_frame in winning_frames:
|
||||
await self.push_frame(winning_frame, direction)
|
||||
def set_llm_name(self, name: str):
|
||||
"""Set the name of the LLM this processor instance is handling."""
|
||||
self._current_llm_name = name
|
||||
|
||||
async def process_frame(self, frame: Frame, direction: FrameDirection):
|
||||
if isinstance(frame, LLMTextFrame):
|
||||
if not LLMRaceProcessor._response_started:
|
||||
# First response wins the race
|
||||
LLMRaceProcessor._winning_llm_name = self._current_llm_name
|
||||
LLMRaceProcessor._response_started = True
|
||||
logger.info(
|
||||
f"🏆 [LLM_RACE] {self._current_llm_name} wins the race! Text: '{frame.text}'"
|
||||
)
|
||||
await self.push_frame(frame, direction)
|
||||
elif LLMRaceProcessor._winning_llm_name == self._current_llm_name:
|
||||
# Continue allowing frames from winning LLM
|
||||
logger.info(f"✅ [LLM_RACE] {self._current_llm_name} continuing: '{frame.text}'")
|
||||
await self.push_frame(frame, direction)
|
||||
else:
|
||||
# Drop frames from losing LLM
|
||||
logger.info(
|
||||
f"❌ [LLM_RACE] Dropping '{frame.text}' from losing LLM: {self._current_llm_name}"
|
||||
)
|
||||
else:
|
||||
# Pass through non-LLM frames
|
||||
# Always pass through non-LLM frames
|
||||
await self.push_frame(frame, direction)
|
||||
|
||||
|
||||
# Class variables to share state between instances
|
||||
LLMRaceProcessor._winning_llm_name = None
|
||||
LLMRaceProcessor._response_started = False
|
||||
|
||||
|
||||
async def run_bot(transport: BaseTransport, runner_args: RunnerArguments):
|
||||
logger.info(f"Starting bot with parallel LLM racing")
|
||||
|
||||
@@ -154,19 +121,35 @@ async def run_bot(transport: BaseTransport, runner_args: RunnerArguments):
|
||||
context = OpenAILLMContext(messages)
|
||||
context_aggregator = llm1.create_context_aggregator(context)
|
||||
|
||||
# Create a second context aggregator that shares the same context
|
||||
context_aggregator2 = llm2.create_context_aggregator(context)
|
||||
# Make sure both LLMs share the same context - they should both process context frames
|
||||
# In a ParallelPipeline, the context frames will be duplicated to both branches
|
||||
|
||||
# Create race processor with both LLMs
|
||||
race_processor = LLMRaceProcessor(llm1, llm2)
|
||||
# Create separate race processors for each LLM to track which one responds first
|
||||
race_processor1 = LLMRaceProcessor()
|
||||
race_processor1.set_llm_name("LLM1")
|
||||
|
||||
# Simple pipeline - the race processor handles the parallel LLM execution internally
|
||||
race_processor2 = LLMRaceProcessor()
|
||||
race_processor2.set_llm_name("LLM2")
|
||||
|
||||
# Create parallel LLM branches using ParallelPipeline
|
||||
parallel_llms = ParallelPipeline(
|
||||
[llm1, race_processor1], # Branch 1: LLM1 -> race processor 1
|
||||
[llm2, race_processor2], # Branch 2: LLM2 -> race processor 2
|
||||
)
|
||||
|
||||
# Set up debug observers with filtering - only log LLM frames going to TTS
|
||||
debug_observer = DebugLogObserver(
|
||||
frame_types={LLMTextFrame: (CartesiaTTSService, FrameEndpoint.DESTINATION)}
|
||||
)
|
||||
llm_observer = LLMLogObserver()
|
||||
|
||||
# Simple pipeline with parallel LLM processing
|
||||
pipeline = Pipeline(
|
||||
[
|
||||
transport.input(), # Transport user input
|
||||
stt, # Speech to text
|
||||
context_aggregator.user(), # User responses (creates context frames for LLMs)
|
||||
race_processor, # Parallel LLM racing processor
|
||||
parallel_llms, # Parallel LLM processing
|
||||
tts, # TTS
|
||||
transport.output(), # Transport bot output
|
||||
context_aggregator.assistant(), # Assistant spoken responses
|
||||
@@ -179,15 +162,18 @@ async def run_bot(transport: BaseTransport, runner_args: RunnerArguments):
|
||||
enable_metrics=True,
|
||||
enable_usage_metrics=True,
|
||||
),
|
||||
# observers=[debug_observer, llm_observer],
|
||||
idle_timeout_secs=runner_args.pipeline_idle_timeout_secs,
|
||||
)
|
||||
|
||||
@transport.event_handler("on_client_connected")
|
||||
async def on_client_connected(transport, client):
|
||||
logger.info(f"Client connected")
|
||||
# Kick off the conversation.
|
||||
# Use a simpler approach - add message to context and push a context frame
|
||||
messages.append({"role": "system", "content": "Please introduce yourself to the user."})
|
||||
await task.queue_frames([context_aggregator.user().get_context_frame()])
|
||||
# Create a new context with the updated messages
|
||||
updated_context = OpenAILLMContext(messages)
|
||||
await task.queue_frames([OpenAILLMContextFrame(context=updated_context)])
|
||||
|
||||
@transport.event_handler("on_client_disconnected")
|
||||
async def on_client_disconnected(transport, client):
|
||||
|
||||
Reference in New Issue
Block a user