GenAIExamples/VideoQnA/videoqna.py

# Copyright (C) 2024 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

import os

from comps import MegaServiceEndpoint, MicroService, ServiceOrchestrator, ServiceRoleType, ServiceType
from comps.cores.mega.utils import handle_message
from comps.cores.proto.api_protocol import (
    ChatCompletionRequest,
    ChatCompletionResponse,
    ChatCompletionResponseChoice,
    ChatMessage,
    UsageInfo,
)
from comps.cores.proto.docarray import LLMParams, TextDoc
from fastapi import Request
from fastapi.responses import StreamingResponse

MEGA_SERVICE_PORT = int(os.getenv("BACKEND_PORT", 8888))
EMBEDDING_SERVICE_HOST_IP = os.getenv("EMBEDDING_SERVICE_HOST_IP", "0.0.0.0")
EMBEDDING_SERVICE_PORT = int(os.getenv("EMBEDDER_PORT", 6990))
RETRIEVER_SERVICE_HOST_IP = os.getenv("RETRIEVER_SERVICE_HOST_IP", "0.0.0.0")
RETRIEVER_SERVICE_PORT = int(os.getenv("RETRIEVER_PORT", 7000))
RERANK_SERVICE_HOST_IP = os.getenv("RERANK_SERVICE_HOST_IP", "0.0.0.0")
RERANK_SERVICE_PORT = int(os.getenv("RERANK_SERVICE_PORT", 8000))
LVM_SERVICE_HOST_IP = os.getenv("LVM_SERVICE_HOST_IP", "0.0.0.0")
LVM_SERVICE_PORT = int(os.getenv("LVM_PORT", 9399))


def align_inputs(self, inputs, cur_node, runtime_graph, llm_parameters_dict, **kwargs):
    service_type = self.services[cur_node].service_type
    if service_type == ServiceType.EMBEDDING:
        if "input" in inputs:
            input_text = inputs["input"]["text"] if isinstance(inputs["input"], dict) else inputs["input"]
            inputs = TextDoc(text=input_text).model_dump()
    return inputs


def align_outputs(self, data, cur_node, inputs, runtime_graph, llm_parameters_dict, **kwargs):
    if self.services[cur_node].service_type == ServiceType.EMBEDDING:
        return {"text": inputs["text"], "embedding": data["embedding"]}
    else:
        return data


class VideoQnAService:
    def __init__(self, host="0.0.0.0", port=8888):
        self.host = host
        self.port = port
        ServiceOrchestrator.align_inputs = align_inputs
        ServiceOrchestrator.align_outputs = align_outputs
        self.megaservice = ServiceOrchestrator()
        self.endpoint = str(MegaServiceEndpoint.VIDEO_RAG_QNA)

    def add_remote_service(self):
        embedding = MicroService(
            name="embedding",
            host=EMBEDDING_SERVICE_HOST_IP,
            port=EMBEDDING_SERVICE_PORT,
            endpoint="/v1/embeddings",
            use_remote_service=True,
            service_type=ServiceType.EMBEDDING,
        )
        retriever = MicroService(
            name="retriever",
            host=RETRIEVER_SERVICE_HOST_IP,
            port=RETRIEVER_SERVICE_PORT,
            endpoint="/v1/retrieval",
            use_remote_service=True,
            service_type=ServiceType.RETRIEVER,
        )
        rerank = MicroService(
            name="rerank",
            host=RERANK_SERVICE_HOST_IP,
            port=RERANK_SERVICE_PORT,
            endpoint="/v1/reranking",
            use_remote_service=True,
            service_type=ServiceType.RERANK,
        )
        lvm = MicroService(
            name="lvm",
            host=LVM_SERVICE_HOST_IP,
            port=LVM_SERVICE_PORT,
            endpoint="/v1/lvm",
            use_remote_service=True,
            service_type=ServiceType.LVM,
        )
        self.megaservice.add(embedding).add(retriever).add(rerank).add(lvm)
        self.megaservice.flow_to(embedding, retriever)
        self.megaservice.flow_to(retriever, rerank)
        self.megaservice.flow_to(rerank, lvm)

    async def handle_request(self, request: Request):
        data = await request.json()
        stream_opt = bool(data.get("stream", False))
        chat_request = ChatCompletionRequest.model_validate(data)
        prompt = handle_message(chat_request.messages)
        parameters = LLMParams(
            max_new_tokens=chat_request.max_tokens if chat_request.max_tokens else 1024,
            top_k=chat_request.top_k if chat_request.top_k else 10,
            top_p=chat_request.top_p if chat_request.top_p else 0.95,
            temperature=chat_request.temperature if chat_request.temperature else 0.01,
            frequency_penalty=chat_request.frequency_penalty if chat_request.frequency_penalty else 0.0,
            presence_penalty=chat_request.presence_penalty if chat_request.presence_penalty else 0.0,
            repetition_penalty=chat_request.repetition_penalty if chat_request.repetition_penalty else 1.03,
            stream=stream_opt,
            chat_template=chat_request.chat_template if chat_request.chat_template else None,
        )
        result_dict, runtime_graph = await self.megaservice.schedule(
            initial_inputs={"input": prompt}, llm_parameters=parameters
        )
        for node, response in result_dict.items():
            # Here it suppose the last microservice in the megaservice is LVM.
            if (
                isinstance(response, StreamingResponse)
                and node == list(self.megaservice.services.keys())[-1]
                and self.megaservice.services[node].service_type == ServiceType.LVM
            ):
                return response
        last_node = runtime_graph.all_leaves()[-1]
        response = result_dict[last_node]["text"]
        choices = []
        usage = UsageInfo()
        choices.append(
            ChatCompletionResponseChoice(
                index=0,
                message=ChatMessage(role="assistant", content=response),
                finish_reason="stop",
            )
        )
        return ChatCompletionResponse(model="videoqna", choices=choices, usage=usage)

    def start(self):
        self.service = MicroService(
            self.__class__.__name__,
            service_role=ServiceRoleType.MEGASERVICE,
            host=self.host,
            port=self.port,
            endpoint=self.endpoint,
            input_datatype=ChatCompletionRequest,
            output_datatype=ChatCompletionResponse,
        )
        self.service.add_route(self.endpoint, self.handle_request, methods=["POST"])
        self.service.start()


if __name__ == "__main__":
    videoqna = VideoQnAService(port=MEGA_SERVICE_PORT)
    videoqna.add_remote_service()
    videoqna.start()