ai-video-fullstack/backend/services/pipecat/pipeline.py

"""管线核心:给定一个 transport + 配置,跑完整的语音闭环。

关键设计:**transport 由调用方传入**,管线本身不关心是 WebRTC 还是 WS。
这就是"同时支持多种输出"的落点——加输出方式不用动这里。

对应 dograh 的 pipeline_builder.py + run_pipeline.py(已砍掉 workflow 引擎/DB/录音/指标)。
"""

from uuid import uuid4

from loguru import logger
from models import AssistantConfig
from services.pipecat.service_factory import create_realtime_service, create_services

from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.frames.frames import (
    EndFrame,
    InputTransportMessageFrame,
    InterruptionFrame,
    LLMFullResponseEndFrame,
    LLMFullResponseStartFrame,
    LLMTextFrame,
    LLMMessagesAppendFrame,
    OutputTransportMessageUrgentFrame,
    TextFrame,
    TTSSpeakFrame,
)
from pipecat.pipeline.pipeline import Pipeline
from pipecat.pipeline.worker import PipelineParams, PipelineWorker
from pipecat.processors.aggregators.llm_context import LLMContext
from pipecat.processors.aggregators.llm_response_universal import (
    LLMAssistantAggregator,
    LLMUserAggregator,
    LLMUserAggregatorParams,
)
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.turns.user_start import (
    TranscriptionUserTurnStartStrategy,
    VADUserTurnStartStrategy,
)
from pipecat.turns.user_turn_strategies import UserTurnStrategies
from pipecat.utils.time import time_now_iso8601
from pipecat.workers.runner import WorkerRunner


def _text_input(message) -> tuple[str, bool] | None:
    """解析现有 user-text 与 RTVI send-text 两种前端文字消息。"""
    if not isinstance(message, dict):
        return None
    if message.get("type") == "user-text":
        text = str(message.get("text") or "").strip()
        return (text, True) if text else None
    if message.get("type") == "send-text":
        data = message.get("data")
        if not isinstance(data, dict):
            return None
        text = str(data.get("content") or "").strip()
        options = data.get("options")
        run_immediately = not isinstance(options, dict) or options.get(
            "run_immediately", True
        )
        return (text, bool(run_immediately)) if text else None
    return None


class TextInputProcessor(FrameProcessor):
    """把 transport 文字消息转换成 LLM 可消费的帧。

    run_immediately(默认/打断):先通过 on_text_input 事件把用户文字交给
    run_pipeline 登记,再用 broadcast_interruption() 打断当前播报。新的 LLM
    回复由 assistant aggregator 确认处理完 interruption 后触发。
    run_immediately=False(RTVI send-text 静默追加):仅把文字写进上下文,
    不打断、不触发推理。
    """

    def __init__(self):
        super().__init__()
        # 立即触发的文字(含打断语义)走 on_text_input;静默追加另走一条事件
        self._register_event_handler("on_text_input")
        self._register_event_handler("on_text_append")
        self._register_event_handler("on_client_ready")

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)

        if not isinstance(frame, InputTransportMessageFrame):
            await self.push_frame(frame, direction)
            return

        if isinstance(frame.message, dict) and frame.message.get("type") == "client-ready":
            await self._call_event_handler("on_client_ready")
            return

        parsed = _text_input(frame.message)
        if not parsed:
            await self.push_frame(frame, direction)
            return

        text, run_immediately = parsed
        if run_immediately:
            # 先登记文字再打断。下一轮 LLM 由 assistant aggregator 在真正处理完
            # InterruptionFrame 后触发,避免新回复被这次 interruption 一起取消。
            await self._call_event_handler("on_text_input", text)
            await self.broadcast_interruption()
        else:
            await self._call_event_handler("on_text_append", text)


class RealtimeTextInputProcessor(FrameProcessor):
    """Route text input directly to a realtime service without cascade semantics."""

    def __init__(self):
        super().__init__()
        self._register_event_handler("on_text_input")
        self._register_event_handler("on_text_append")

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)

        if not isinstance(frame, InputTransportMessageFrame):
            await self.push_frame(frame, direction)
            return

        parsed = _text_input(frame.message)
        if not parsed:
            await self.push_frame(frame, direction)
            return

        text, run_immediately = parsed
        await self._call_event_handler(
            "on_text_input" if run_immediately else "on_text_append",
            text,
        )


class PassthroughLLMAssistantAggregator(LLMAssistantAggregator):
    """聚合 LLM 回复进上下文,同时继续把回复帧交给下游 TTS。"""

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self._register_event_handler("on_interruption_processed")
        self._register_event_handler("on_assistant_text_start")
        self._register_event_handler("on_assistant_text_delta")
        self._register_event_handler("on_assistant_text_end")
        self._stream_turn_id: str | None = None
        self._stream_timestamp = ""
        self._stream_text = ""

    async def process_frame(self, frame, direction: FrameDirection):
        await super().process_frame(frame, direction)

        if isinstance(frame, LLMFullResponseStartFrame):
            self._stream_turn_id = uuid4().hex
            self._stream_timestamp = time_now_iso8601()
            self._stream_text = ""
            await self._call_event_handler(
                "on_assistant_text_start",
                self._stream_turn_id,
                self._stream_timestamp,
            )
        elif isinstance(frame, LLMTextFrame) and self._stream_turn_id:
            self._stream_text += frame.text
            await self._call_event_handler(
                "on_assistant_text_delta",
                self._stream_turn_id,
                frame.text,
            )
        elif isinstance(frame, LLMFullResponseEndFrame):
            await self._finish_text_stream(interrupted=False)

        # LLMAssistantAggregator 默认会消费这些帧。放在 TTS 前用于中断时保存
        # 已生成前缀时,必须显式透传,否则 TTS 收不到任何 LLM 回复。
        if isinstance(
            frame,
            (LLMFullResponseStartFrame, LLMFullResponseEndFrame, TextFrame),
        ):
            await self.push_frame(frame, direction)
        elif isinstance(frame, InterruptionFrame):
            await self._finish_text_stream(interrupted=True)
            await self._call_event_handler("on_interruption_processed")

    async def _finish_text_stream(self, *, interrupted: bool):
        if not self._stream_turn_id:
            return
        await self._call_event_handler(
            "on_assistant_text_end",
            self._stream_turn_id,
            self._stream_text,
            interrupted,
        )
        self._stream_turn_id = None
        self._stream_timestamp = ""
        self._stream_text = ""


async def run_pipeline(transport, cfg: AssistantConfig) -> None:
    """在给定 transport 上构建并运行管线,直到连接结束。

    Args:
        transport: 任意 pipecat transport(WebRTC / WS / 电话…),
                   只要有 .input() / .output() / event_handler 即可。
        cfg:       助手配置(随请求内联传入)。
    """
    logger.info(f"启动管线: assistant={cfg.name} mode={cfg.runtimeMode}")

    if cfg.runtimeMode == "realtime":
        await run_realtime_pipeline(transport, cfg)
        return

    stt, llm, tts = create_services(cfg)

    context = LLMContext(messages=[{"role": "system", "content": cfg.prompt}])
    user_aggregator = LLMUserAggregator(
        context,
        params=LLMUserAggregatorParams(
            vad_analyzer=SileroVADAnalyzer(),
            user_turn_strategies=UserTurnStrategies(
                start=[
                    VADUserTurnStartStrategy(enable_interruptions=cfg.enableInterrupt),
                    TranscriptionUserTurnStartStrategy(
                        enable_interruptions=cfg.enableInterrupt
                    ),
                ]
            ),
        ),
    )
    assistant_aggregator = PassthroughLLMAssistantAggregator(context)
    text_input = TextInputProcessor()

    pipeline = Pipeline(
        [
            transport.input(),
            text_input,
            stt,
            user_aggregator,
            llm,
            # Aggregate the streamed LLM text before TTS. On interruption,
            # Pipecat commits the generated prefix immediately instead of
            # waiting for a TTS provider to emit spoken-text/timestamp frames.
            assistant_aggregator,
            tts,
            transport.output(),
        ]
    )

    worker = PipelineWorker(
        pipeline,
        params=PipelineParams(
            enable_metrics=False,
        ),
        enable_rtvi=False,
    )

    async def queue_transcript(role: str, content: str, timestamp: str) -> None:
        if content:
            await worker.queue_frame(
                OutputTransportMessageUrgentFrame(
                    message={
                        "type": "transcript",
                        "role": role,
                        "content": content,
                        "timestamp": timestamp,
                    },
                )
            )

    greeting_transcript_sent = False
    pending_text_inputs: list[str] = []

    async def append_user_text_to_context(text: str, *, run_llm: bool) -> None:
        await worker.queue_frame(
            LLMMessagesAppendFrame(
                messages=[{"role": "user", "content": text}],
                run_llm=run_llm,
            )
        )

    @user_aggregator.event_handler("on_user_turn_stopped")
    async def on_user_turn_stopped(_aggregator, _strategy, message):
        await queue_transcript("user", message.content, message.timestamp)

    @assistant_aggregator.event_handler("on_assistant_text_start")
    async def on_assistant_text_start(_aggregator, turn_id, timestamp):
        await worker.queue_frame(
            OutputTransportMessageUrgentFrame(
                message={
                    "type": "assistant-text-start",
                    "turn_id": turn_id,
                    "timestamp": timestamp,
                }
            )
        )

    @assistant_aggregator.event_handler("on_assistant_text_delta")
    async def on_assistant_text_delta(_aggregator, turn_id, delta):
        await worker.queue_frame(
            OutputTransportMessageUrgentFrame(
                message={
                    "type": "assistant-text-delta",
                    "turn_id": turn_id,
                    "delta": delta,
                }
            )
        )

    @assistant_aggregator.event_handler("on_assistant_text_end")
    async def on_assistant_text_end(_aggregator, turn_id, content, interrupted):
        await worker.queue_frame(
            OutputTransportMessageUrgentFrame(
                message={
                    "type": "assistant-text-end",
                    "turn_id": turn_id,
                    "content": content,
                    "interrupted": interrupted,
                }
            )
        )

    @text_input.event_handler("on_text_input")
    async def on_text_input(_processor, text):
        pending_text_inputs.append(text)
        # 前端显示不依赖 interruption 后续事件,必须在打断前先排入发送队列。
        await queue_transcript("user", text, time_now_iso8601())

    @assistant_aggregator.event_handler("on_interruption_processed")
    async def on_interruption_processed(_aggregator):
        if not pending_text_inputs:
            return
        text = pending_text_inputs.pop(0)
        # assistant aggregator 已处理完 interruption,现在再启动下一轮 LLM。
        await append_user_text_to_context(text, run_llm=True)

    @text_input.event_handler("on_text_append")
    async def on_text_append(_processor, text):
        # 静默追加:写进上下文但不打断、不触发推理;transcript 照常上报
        await queue_transcript("user", text, time_now_iso8601())
        await append_user_text_to_context(text, run_llm=False)

    @text_input.event_handler("on_client_ready")
    async def on_client_ready(_processor):
        nonlocal greeting_transcript_sent
        if cfg.greeting and not greeting_transcript_sent:
            greeting_transcript_sent = True
            await queue_transcript("assistant", cfg.greeting, time_now_iso8601())

    @transport.event_handler("on_client_connected")
    async def on_client_connected(_transport, _client):
        if cfg.greeting:
            context.add_message({"role": "assistant", "content": cfg.greeting})
            await worker.queue_frame(TTSSpeakFrame(cfg.greeting, append_to_context=False))

    @transport.event_handler("on_client_disconnected")
    async def on_client_disconnected(_transport, _client):
        logger.info("对端断开,结束管线")
        await worker.queue_frame(EndFrame())

    runner = WorkerRunner(handle_sigint=False)
    await runner.add_workers(worker)
    await runner.run()
    logger.info("管线已结束")


async def run_realtime_pipeline(transport, cfg: AssistantConfig) -> None:
    """Run a speech-to-speech model that owns ASR, reasoning, and synthesis."""
    realtime = create_realtime_service(cfg)
    text_input = RealtimeTextInputProcessor()

    pipeline = Pipeline(
        [
            transport.input(),
            text_input,
            realtime,
            transport.output(),
        ]
    )
    worker = PipelineWorker(
        pipeline,
        params=PipelineParams(
            enable_metrics=False,
            audio_in_sample_rate=int(
                cfg.realtime_values.get("inputSampleRate") or 24000
            ),
            audio_out_sample_rate=int(
                cfg.realtime_values.get("outputSampleRate") or 24000
            ),
        ),
        enable_rtvi=False,
    )

    async def queue_transcript(role: str, content: str) -> None:
        if content:
            await worker.queue_frame(
                OutputTransportMessageUrgentFrame(
                    message={
                        "type": "transcript",
                        "role": role,
                        "content": content,
                        "timestamp": time_now_iso8601(),
                    },
                )
            )

    @text_input.event_handler("on_text_input")
    async def on_text_input(_processor, text):
        await queue_transcript("user", text)
        await realtime.interrupt()
        await realtime.send_text(text, run_immediately=True)

    @text_input.event_handler("on_text_append")
    async def on_text_append(_processor, text):
        await queue_transcript("user", text)
        await realtime.send_text(text, run_immediately=False)

    @transport.event_handler("on_client_connected")
    async def on_client_connected(_transport, _client):
        if cfg.greeting:
            await realtime.speak(cfg.greeting)

    @transport.event_handler("on_client_disconnected")
    async def on_client_disconnected(_transport, _client):
        logger.info("Realtime 对端断开,结束管线")
        await worker.queue_frame(EndFrame())

    runner = WorkerRunner(handle_sigint=False)
    await runner.add_workers(worker)
    await runner.run()
    logger.info("Realtime 管线已结束")