routes/llm.py

import json
import logging
import os
from typing import Annotated, AsyncGenerator, Optional
from uuid import UUID

from components.services.dialogue import DialogueService
from fastapi.responses import StreamingResponse

from components.services.dataset import DatasetService
from components.services.entity import EntityService
from fastapi import APIRouter, Depends, HTTPException

import common.dependencies as DI
from common.configuration import Configuration, Query
from components.llm.common import ChatRequest, LlmParams, LlmPredictParams, Message
from components.llm.deepinfra_api import DeepInfraApi
from components.llm.utils import append_llm_response_to_history
from components.services.llm_config import LLMConfigService
from components.services.llm_prompt import LlmPromptService

router = APIRouter(prefix='/llm', tags=['LLM chat'])
logger = logging.getLogger(__name__)

conf = DI.get_config()
llm_params = LlmParams(
    **{
        "url": conf.llm_config.base_url,
        "model": conf.llm_config.model,
        "tokenizer": "unsloth/Llama-3.3-70B-Instruct",
        "type": "deepinfra",
        "default": True,
        "predict_params": LlmPredictParams(
            temperature=0.15,
            top_p=0.95,
            min_p=0.05,
            seed=42,
            repetition_penalty=1.2,
            presence_penalty=1.1,
            n_predict=2000,
        ),
        "api_key": os.environ.get(conf.llm_config.api_key_env),
        "context_length": 128000,
    }
)
# TODO: унести в DI
llm_api = DeepInfraApi(params=llm_params)

# TODO: Вынести
def get_last_user_message(chat_request: ChatRequest) -> Optional[Message]:
    return next(
        (
            msg
            for msg in reversed(chat_request.history)
            if msg.role == "user"
            and (msg.searchResults is None or not msg.searchResults)
        ),
        None,
    )

def insert_search_results_to_message(
    chat_request: ChatRequest, new_content: str
) -> bool:
    for msg in reversed(chat_request.history):
        if msg.role == "user" and (
            msg.searchResults is None or not msg.searchResults
        ):
            msg.content = new_content
            return True
    return False

def try_insert_search_results(
    chat_request: ChatRequest, search_results: str
) -> bool:
    for msg in reversed(chat_request.history):
        if msg.role == "user" and not msg.searchResults:
            msg.searchResults = search_results
            return True
    return False

def collapse_history_to_first_message(chat_request: ChatRequest) -> ChatRequest:
    """
    Сворачивает историю в первое сообщение и возвращает новый объект ChatRequest.
    Формат:
    <search-results>[Источник] - текст</search-results>
    role: текст сообщения
    """
    if not chat_request.history:
        return ChatRequest(history=[])

    # Собираем историю в одну строку
    collapsed_content = []
    for msg in chat_request.history:
        # Добавляем search-results, если они есть
        if msg.searchResults:
            collapsed_content.append(f"<search-results>{msg.searchResults}</search-results>")
        # Добавляем текст сообщения с указанием роли
        if msg.content.strip():
            collapsed_content.append(f"{msg.role}: {msg.content.strip()}")

    # Формируем финальный текст с переносами строк
    new_content = "\n".join(collapsed_content)

    # Создаем новое сообщение и новый объект ChatRequest
    new_message = Message(
        role='user',
        content=new_content,
        searchResults=''
    )
    return ChatRequest(history=[new_message])
        
async def sse_generator(request: ChatRequest, llm_api: DeepInfraApi, system_prompt: str, 
                        predict_params: LlmPredictParams,
                        dataset_service: DatasetService, 
                        entity_service: EntityService,
                        dialogue_service: DialogueService) -> AsyncGenerator[str, None]:
    """
    Генератор для стриминга ответа LLM через SSE.
    """
    
    qe_result = await dialogue_service.get_qe_result(request.history)    
    
    if qe_result.use_search and qe_result.search_query is not None:
        dataset = dataset_service.get_current_dataset()
        if dataset is None:
            raise HTTPException(status_code=400, detail="Dataset not found")
        _, scores, chunk_ids = entity_service.search_similar(qe_result.search_query, dataset.id)
        chunks = entity_service.chunk_repository.get_chunks_by_ids(chunk_ids)
        text_chunks = entity_service.build_text(chunks, scores)
        search_results_event = {
            "event": "search_results",
            "data": f"{text_chunks}"
        }
        yield f"data: {json.dumps(search_results_event, ensure_ascii=False)}\n\n"

        # new_message = f'<search-results>\n{text_chunks}\n</search-results>\n{last_query.content}'
        
        try_insert_search_results(request, text_chunks)
        
        
    # Сворачиваем историю в первое сообщение
    collapsed_request = collapse_history_to_first_message(request)
                
    # Стриминг токенов ответа
    async for token in llm_api.get_predict_chat_generator(collapsed_request, system_prompt, predict_params):
        token_event = {"event": "token", "data": token}
        # logger.info(f"Streaming token: {token}")
        yield f"data: {json.dumps(token_event, ensure_ascii=False)}\n\n"

    # Финальное событие
    yield "data: {\"event\": \"done\"}\n\n"
    

@router.post("/chat/stream")
async def chat_stream(
    request: ChatRequest,
    config: Annotated[Configuration, Depends(DI.get_config)],
    llm_api: Annotated[DeepInfraApi, Depends(DI.get_llm_service)],
    prompt_service: Annotated[LlmPromptService, Depends(DI.get_llm_prompt_service)],
    llm_config_service: Annotated[LLMConfigService, Depends(DI.get_llm_config_service)],
    entity_service: Annotated[EntityService, Depends(DI.get_entity_service)],
    dataset_service: Annotated[DatasetService, Depends(DI.get_dataset_service)],
    dialogue_service: Annotated[DialogueService, Depends(DI.get_dialogue_service)],
):
    try:
        p = llm_config_service.get_default()
        system_prompt = prompt_service.get_default()

        predict_params = LlmPredictParams(
            temperature=p.temperature,
            top_p=p.top_p,
            min_p=p.min_p,
            seed=p.seed,
            frequency_penalty=p.frequency_penalty,
            presence_penalty=p.presence_penalty,
            n_predict=p.n_predict,
            stop=[],
        )

        headers = {
            "Content-Type": "text/event-stream",
            "Cache-Control": "no-cache",
            "Connection": "keep-alive",
            "Access-Control-Allow-Origin": "*",
        }
        return StreamingResponse(
            sse_generator(request, llm_api, system_prompt.text, predict_params, dataset_service, entity_service, dialogue_service),
            media_type="text/event-stream",
            headers=headers
        )
    except Exception as e:
        logger.error(f"Error in SSE chat stream: {str(e)}", stack_info=True)
        raise HTTPException(status_code=500, detail=str(e))

@router.post("/chat")
async def chat(
    request: ChatRequest,
    config: Annotated[Configuration, Depends(DI.get_config)],
    llm_api: Annotated[DeepInfraApi, Depends(DI.get_llm_service)],
    prompt_service: Annotated[LlmPromptService, Depends(DI.get_llm_prompt_service)],
    llm_config_service: Annotated[LLMConfigService, Depends(DI.get_llm_config_service)],
    entity_service: Annotated[EntityService, Depends(DI.get_entity_service)],
    dataset_service: Annotated[DatasetService, Depends(DI.get_dataset_service)],
    dialogue_service: Annotated[DialogueService, Depends(DI.get_dialogue_service)],
):
    try:
        p = llm_config_service.get_default()
        system_prompt = prompt_service.get_default()

        predict_params = LlmPredictParams(
            temperature=p.temperature,
            top_p=p.top_p,
            min_p=p.min_p,
            seed=p.seed,
            frequency_penalty=p.frequency_penalty,
            presence_penalty=p.presence_penalty,
            n_predict=p.n_predict,
            stop=[],
        )

        qe_result = await dialogue_service.get_qe_result(request.history)
        last_message = get_last_user_message(request)

        logger.info(f"qe_result: {qe_result}")

        if qe_result.use_search and qe_result.search_query is not None:
            dataset = dataset_service.get_current_dataset()
            if dataset is None:
                raise HTTPException(status_code=400, detail="Dataset not found")
            logger.info(f"qe_result.search_query: {qe_result.search_query}")
            _, scores, chunk_ids = entity_service.search_similar(qe_result.search_query, dataset.id)
            
            chunks = entity_service.chunk_repository.get_chunks_by_ids(chunk_ids)
            
            logger.info(f"chunk_ids: {chunk_ids[:3]}...{chunk_ids[-3:]}")
            logger.info(f"scores: {scores[:3]}...{scores[-3:]}")
            
            text_chunks = entity_service.build_text(chunks, scores)
            
            logger.info(f"text_chunks: {text_chunks[:3]}...{text_chunks[-3:]}")

            new_message = f'{last_message.content} /n<search-results>/n{text_chunks}/n</search-results>'
            insert_search_results_to_message(request, new_message)
            
        logger.info(f"request: {request}")

        response = await llm_api.predict_chat_stream(
            request, system_prompt.text, predict_params
        )
        result = append_llm_response_to_history(request, response)
        return result
    except Exception as e:
        logger.error(
            f"Error processing LLM request: {str(e)}", stack_info=True, stacklevel=10
        )
        return {"error": str(e)}