333 lines
12 KiB
Python
333 lines
12 KiB
Python
import asyncio
|
|
import re
|
|
|
|
from dailyai.pipeline.frame_processor import FrameProcessor
|
|
|
|
from dailyai.pipeline.frames import (
|
|
EndFrame,
|
|
EndPipeFrame,
|
|
Frame,
|
|
ImageFrame,
|
|
LLMMessagesQueueFrame,
|
|
LLMResponseEndFrame,
|
|
LLMResponseStartFrame,
|
|
TextFrame,
|
|
TranscriptionQueueFrame,
|
|
)
|
|
from dailyai.pipeline.pipeline import Pipeline
|
|
from dailyai.services.ai_services import AIService
|
|
|
|
from typing import AsyncGenerator, Coroutine, List
|
|
|
|
class LLMResponseAggregator(FrameProcessor):
|
|
def __init__(self, messages: list[dict]):
|
|
self.aggregation = ""
|
|
self.aggregating = False
|
|
self.messages = messages
|
|
|
|
async def process_frame(
|
|
self, frame: Frame
|
|
) -> AsyncGenerator[Frame, None]:
|
|
if isinstance(frame, LLMResponseStartFrame):
|
|
self.aggregating = True
|
|
elif isinstance(frame, LLMResponseEndFrame):
|
|
self.aggregating = False
|
|
self.messages.append({"role": "assistant", "content": self.aggregation})
|
|
self.aggregation = ""
|
|
yield LLMMessagesQueueFrame(self.messages)
|
|
elif isinstance(frame, TextFrame) and self.aggregating:
|
|
self.aggregation += frame.text
|
|
yield frame
|
|
else:
|
|
yield frame
|
|
|
|
|
|
class LLMContextAggregator(AIService):
|
|
def __init__(
|
|
self,
|
|
messages: list[dict],
|
|
role: str,
|
|
bot_participant_id=None,
|
|
complete_sentences=True,
|
|
pass_through=True,
|
|
):
|
|
super().__init__()
|
|
self.messages = messages
|
|
self.bot_participant_id = bot_participant_id
|
|
self.role = role
|
|
self.sentence = ""
|
|
self.complete_sentences = complete_sentences
|
|
self.pass_through = pass_through
|
|
|
|
async def process_frame(
|
|
self, frame: Frame
|
|
) -> AsyncGenerator[Frame, None]:
|
|
# We don't do anything with non-text frames, pass it along to next in the pipeline.
|
|
if not isinstance(frame, TextFrame):
|
|
yield frame
|
|
return
|
|
|
|
# Ignore transcription frames from the bot
|
|
if isinstance(frame, TranscriptionQueueFrame):
|
|
if frame.participantId == self.bot_participant_id:
|
|
return
|
|
|
|
# The common case for "pass through" is receiving frames from the LLM that we'll
|
|
# use to update the "assistant" LLM messages, but also passing the text frames
|
|
# along to a TTS service to be spoken to the user.
|
|
if self.pass_through:
|
|
yield frame
|
|
|
|
# TODO: split up transcription by participant
|
|
if self.complete_sentences:
|
|
# type: ignore -- the linter thinks this isn't a TextQueueFrame, even
|
|
# though we check it above
|
|
self.sentence += frame.text
|
|
if self.sentence.endswith((".", "?", "!")):
|
|
self.messages.append({"role": self.role, "content": self.sentence})
|
|
self.sentence = ""
|
|
yield LLMMessagesQueueFrame(self.messages)
|
|
else:
|
|
# type: ignore -- the linter thinks this isn't a TextQueueFrame, even
|
|
# though we check it above
|
|
self.messages.append({"role": self.role, "content": frame.text})
|
|
yield LLMMessagesQueueFrame(self.messages)
|
|
|
|
class LLMUserContextAggregator(LLMContextAggregator):
|
|
def __init__(
|
|
self, messages: list[dict], bot_participant_id=None, complete_sentences=True
|
|
):
|
|
super().__init__(
|
|
messages, "user", bot_participant_id, complete_sentences, pass_through=False
|
|
)
|
|
|
|
|
|
class LLMAssistantContextAggregator(LLMContextAggregator):
|
|
def __init__(
|
|
self, messages: list[dict], bot_participant_id=None, complete_sentences=True
|
|
):
|
|
super().__init__(
|
|
messages,
|
|
"assistant",
|
|
bot_participant_id,
|
|
complete_sentences,
|
|
pass_through=True,
|
|
)
|
|
|
|
|
|
class SentenceAggregator(FrameProcessor):
|
|
"""This frame processor aggregates text frames into complete sentences.
|
|
|
|
Frame input/output:
|
|
TextFrame("Hello,") -> None
|
|
TextFrame(" world.") -> TextFrame("Hello world.")
|
|
|
|
Doctest:
|
|
>>> async def print_frames(aggregator, frame):
|
|
... async for frame in aggregator.process_frame(frame):
|
|
... print(frame.text)
|
|
|
|
>>> aggregator = SentenceAggregator()
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame("Hello,")))
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame(" world.")))
|
|
Hello, world.
|
|
"""
|
|
def __init__(self):
|
|
self.aggregation = ""
|
|
|
|
async def process_frame(
|
|
self, frame: Frame
|
|
) -> AsyncGenerator[Frame, None]:
|
|
if isinstance(frame, TextFrame):
|
|
m = re.search("(.*[?.!])(.*)", frame.text)
|
|
if m:
|
|
yield TextFrame(self.aggregation + m.group(1))
|
|
self.aggregation = m.group(2)
|
|
else:
|
|
self.aggregation += frame.text
|
|
elif isinstance(frame, EndFrame):
|
|
if self.aggregation:
|
|
yield TextFrame(self.aggregation)
|
|
yield frame
|
|
else:
|
|
yield frame
|
|
|
|
|
|
class LLMFullResponseAggregator(FrameProcessor):
|
|
"""This class aggregates Text frames until it receives a
|
|
LLMResponseEndFrame, then emits the concatenated text as
|
|
a single text frame.
|
|
|
|
given the following frames:
|
|
|
|
TextFrame("Hello,")
|
|
TextFrame(" world.")
|
|
TextFrame(" I am")
|
|
TextFrame(" an LLM.")
|
|
LLMResponseEndFrame()]
|
|
|
|
this processor will yield nothing for the first 4 frames, then
|
|
|
|
TextFrame("Hello, world. I am an LLM.")
|
|
LLMResponseEndFrame()
|
|
|
|
when passed the last frame.
|
|
|
|
>>> async def print_frames(aggregator, frame):
|
|
... async for frame in aggregator.process_frame(frame):
|
|
... if isinstance(frame, TextFrame):
|
|
... print(frame.text)
|
|
... else:
|
|
... print(frame.__class__.__name__)
|
|
|
|
>>> aggregator = LLMFullResponseAggregator()
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame("Hello,")))
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame(" world.")))
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame(" I am")))
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame(" an LLM.")))
|
|
>>> asyncio.run(print_frames(aggregator, LLMResponseEndFrame()))
|
|
Hello, world. I am an LLM.
|
|
LLMResponseEndFrame
|
|
"""
|
|
def __init__(self):
|
|
self.aggregation = ""
|
|
|
|
async def process_frame(
|
|
self, frame: Frame
|
|
) -> AsyncGenerator[Frame, None]:
|
|
if isinstance(frame, TextFrame):
|
|
self.aggregation += frame.text
|
|
elif isinstance(frame, LLMResponseEndFrame):
|
|
yield TextFrame(self.aggregation)
|
|
yield frame
|
|
self.aggregation = ""
|
|
else:
|
|
yield frame
|
|
|
|
|
|
class StatelessTextTransformer(FrameProcessor):
|
|
"""This processor calls the given function on any text in a text frame.
|
|
|
|
>>> async def print_frames(aggregator, frame):
|
|
... async for frame in aggregator.process_frame(frame):
|
|
... print(frame.text)
|
|
|
|
>>> aggregator = StatelessTextTransformer(lambda x: x.upper())
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame("Hello")))
|
|
HELLO
|
|
"""
|
|
|
|
def __init__(self, transform_fn):
|
|
self.transform_fn = transform_fn
|
|
|
|
async def process_frame(self, frame: Frame) -> AsyncGenerator[Frame, None]:
|
|
if isinstance(frame, TextFrame):
|
|
result = self.transform_fn(frame.text)
|
|
if isinstance(result, Coroutine):
|
|
result = await result
|
|
|
|
yield TextFrame(result)
|
|
else:
|
|
yield frame
|
|
|
|
class ParallelPipeline(FrameProcessor):
|
|
""" Run multiple pipelines in parallel.
|
|
|
|
This class takes frames from its source queue and sends them to each
|
|
sub-pipeline. Each sub-pipeline emits its frames into this class's
|
|
sink queue. No guarantees are made about the ordering of frames in
|
|
the sink queue (that is, no sub-pipeline has higher priority than
|
|
any other, frames are put on the sink in the order they're emitted
|
|
by the sub-pipelines).
|
|
|
|
After each frame is taken from this class's source queue and placed
|
|
in each sub-pipeline's source queue, an EndPipeFrame is put on each
|
|
sub-pipeline's source queue. This indicates to the sub-pipe runner
|
|
that it should exit.
|
|
|
|
Since frame handlers pass through unhandled frames by convention, this
|
|
class de-dupes frames in its sink before yielding them.
|
|
"""
|
|
def __init__(self, pipeline_definitions: List[List[FrameProcessor]]):
|
|
self.sources = [asyncio.Queue() for _ in pipeline_definitions]
|
|
self.sink: asyncio.Queue[Frame] = asyncio.Queue()
|
|
self.pipelines: list[Pipeline] = [
|
|
Pipeline(
|
|
pipeline_definition,
|
|
source,
|
|
self.sink,
|
|
)
|
|
for source, pipeline_definition in zip(self.sources, pipeline_definitions)
|
|
]
|
|
|
|
async def process_frame(self, frame: Frame) -> AsyncGenerator[Frame, None]:
|
|
for source in self.sources:
|
|
await source.put(frame)
|
|
await source.put(EndPipeFrame())
|
|
|
|
await asyncio.gather(*[pipeline.run_pipeline() for pipeline in self.pipelines])
|
|
|
|
seen_ids = set()
|
|
while not self.sink.empty():
|
|
frame = await self.sink.get()
|
|
|
|
# de-dup frames. Because the convention is to yield a frame that isn't processed,
|
|
# each pipeline will likely yield the same frame, so we will end up with _n_ copies
|
|
# of unprocessed frames where _n_ is the number of parallel pipes that don't
|
|
# process that frame.
|
|
if id(frame) in seen_ids:
|
|
continue
|
|
seen_ids.add(id(frame))
|
|
|
|
# Skip passing along EndParallelPipeQueueFrame, because we use them for our own flow control.
|
|
if not isinstance(frame, EndPipeFrame):
|
|
yield frame
|
|
|
|
class GatedAggregator(FrameProcessor):
|
|
"""Accumulate frames, with custom functions to start and stop accumulation.
|
|
Yields gate-opening frame before any accumulated frames, then ensuing frames
|
|
until and not including the gate-closed frame.
|
|
|
|
>>> async def print_frames(aggregator, frame):
|
|
... async for frame in aggregator.process_frame(frame):
|
|
... if isinstance(frame, TextFrame):
|
|
... print(frame.text)
|
|
... else:
|
|
... print(frame.__class__.__name__)
|
|
|
|
>>> aggregator = GatedAggregator(
|
|
... gate_close_fn=lambda x: isinstance(x, LLMResponseStartFrame),
|
|
... gate_open_fn=lambda x: isinstance(x, ImageFrame),
|
|
... start_open=False)
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame("Hello")))
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame("Hello again.")))
|
|
>>> asyncio.run(print_frames(aggregator, ImageFrame(url='', image=bytes([]))))
|
|
ImageFrame
|
|
Hello
|
|
Hello again.
|
|
>>> asyncio.run(print_frames(aggregator, TextFrame("Goodbye.")))
|
|
Goodbye.
|
|
"""
|
|
def __init__(self, gate_open_fn, gate_close_fn, start_open):
|
|
self.gate_open_fn = gate_open_fn
|
|
self.gate_close_fn = gate_close_fn
|
|
self.gate_open = start_open
|
|
self.accumulator: List[Frame] = []
|
|
|
|
async def process_frame(self, frame: Frame) -> AsyncGenerator[Frame, None]:
|
|
if self.gate_open:
|
|
if self.gate_close_fn(frame):
|
|
self.gate_open = False
|
|
else:
|
|
if self.gate_open_fn(frame):
|
|
self.gate_open = True
|
|
|
|
if self.gate_open:
|
|
yield frame
|
|
if self.accumulator:
|
|
for frame in self.accumulator:
|
|
yield frame
|
|
self.accumulator = []
|
|
else:
|
|
self.accumulator.append(frame)
|