main.py

from fastapi import FastAPI, Request, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
import json
import re
import os
from llm.common import LlmParams, LlmPredictParams
from llm.deepinfra_api import DeepInfraApi
from llm import prompts
from prompts import gettable
from dotenv import load_dotenv
import uvicorn

# Загрузка переменных окружения из файла .env
load_dotenv()

LLM_API_URL = os.getenv("LLM_API_URL", "https://api.deepinfra.com")
LLM_API_KEY = os.getenv("DEEPINFRA_API_KEY", "")
LLM_NAME = os.getenv("LLM_NAME", "meta-llama/Llama-3.3-70B-Instruct-Turbo")
TOKENIZER_NAME = os.getenv("TOKENIZER_NAME", "unsloth/Llama-3.3-70B-Instruct")

default_llm_params = LlmParams(
    url=LLM_API_URL,
    api_key=LLM_API_KEY,
    model=LLM_NAME,
    tokenizer=TOKENIZER_NAME,
    context_length=130000,
    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=6000
    )
)
llm_api = DeepInfraApi(default_llm_params)

app = FastAPI()

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"]
)

class TextRequest(BaseModel):
    text: str
    projects: list[str] = []

async def trim_prompt(prompt: str, system_prompt: str):
    
    result = await llm_api.tokenize(prompt+system_prompt)
    result_system = await llm_api.tokenize(system_prompt)
    
    # в случае ошибки при токенизации, вернем исходную строку безопасной длины
    if result["result"] is None or result_system is None:
        return prompt[llm_api.params.context_length / 3]
    
    #вероятно, часть уходит на форматирование чата, надо проверить
    max_length = result["max_length"] - len(result["result"]) - len(result_system["result"]) - llm_api.params.predict_params.n_predict

    detokenized_str = await llm_api.detokenize(result["result"][:max_length])    
    
    # в случае ошибки при детокенизации, вернем исходную строку безопасной длины
    if detokenized_str is None:
        return prompt[llm_api.params.context_length / 3]

    return detokenized_str
    

def validate_json_format(data):
    """ Проверяет, соответствует ли JSON ожидаемому формату. """
    if not isinstance(data, list):
        return "JSON должен быть списком объектов."

    # Возможно, в дальнейшем стоит описать менее детально, пока так для отладки
    for item in data:
        if not isinstance(item, dict):
            return "Элементы списка должны быть объектами (dict)."
        if "name" not in item or "data" not in item:
            return "Каждый объект должен содержать ключи 'name' и 'data'."
        if not isinstance(item["name"], str):
            return "'name' должен быть строкой."
        if not isinstance(item["data"], dict):
            return "'data' должен быть объектом (dict)."
        if "columns" not in item["data"] or "rows" not in item["data"]:
            return "'data' должен содержать 'columns' и 'rows'."
        if not isinstance(item["data"]["columns"], list) or not all(isinstance(col, str) for col in item["data"]["columns"]):
            return "'columns' должен быть списком строк."
        if not isinstance(item["data"]["rows"], list) or not all(isinstance(row, list) for row in item["data"]["rows"]):
            return "'rows' должен быть списком списков."

    return None  # Ошибок нет        
        
    
@app.post("/extracttable")
async def extracttable_route(request: TextRequest):
    prompt = gettable.USER_PROMPT.format(query=request.text)
    system_prompt=gettable.SYSTEM_PROMPT
    prompt = await trim_prompt(prompt, system_prompt)
    response = await llm_api.predict(prompt, system_prompt=system_prompt)
    
    result = {"response": None, "error": None, "raw": response}
    
    if "JSON" not in response:
        result["error"] = "Строка не содержит 'JSON'"
        return result
    
    prefix, json_str = response.split("JSON", 1)
    json_str = json_str.strip()
    
    if not json_str:
        result["error"] = "После 'JSON' отсутствует JSON"
        return result
    
    try:
        result["response"] = json.loads(json_str)
        result["raw"] = prefix.strip()
        validation_error = validate_json_format(result["response"])
        if validation_error:
            result["error"] = validation_error
        else:
            result["response"] = result["response"]
            result["raw"] = prefix.strip()
    except json.JSONDecodeError as e:
        result["error"] = f"Ошибка декодирования JSON: {e}"
        
    return result

@app.get("/health")
def health():
    return {"status": "ok"}

async def generate_response(prompt):
    prompt = await trim_prompt(prompt)
    return await llm_api.predict(prompt)

@app.post("/getsummary")
async def getsummary_route(request: TextRequest):
    return {"result": await generate_response(prompts.GET_SUMMARY.format(text=request.text))}

@app.post("/cleantext")
async def cleantext_route(request: TextRequest):
    return {"result": await generate_response(prompts.CLEAN_TEXT.format(text=request.text))}

@app.post("/getfollowup")
async def getfollowup_route(request: TextRequest):
    return {"result": await generate_response(prompts.GET_FOLLOWUP.format(text=request.text))}

@app.post("/getagenda")
async def getagenda_route(request: TextRequest):
    return {"result": await generate_response(prompts.GET_AGENDA.format(text=request.text))}

@app.post("/gethighlights")
async def gethighlights_route(request: TextRequest):
    return {"result": await generate_response(prompts.GET_HIGHLIGHTS.format(text=request.text))}

@app.post("/getprojectinfo")
async def getprojectinfo_route(request: TextRequest):
    prompts_list = [
        prompts.GET_PROJECT_INFO_NAMES.format(text=request.text),
        prompts.GET_PROJECT_INFO_AGENDA.format(text=request.text)
    ]
    main_info = "\n\n".join([await generate_response(p) for p in prompts_list])
    return {"result": main_info.strip().replace("Конец ответа", "").replace('</s>', '')}

@app.post("/getprojectlist")
async def getprojectlist_route(request: TextRequest):
    list_of_projects = await generate_response(prompts.GET_PROJECT_LIST.format(text=request.text))
    projects = [f"Проект {x}" for x in list_of_projects.split("Проект ") if x][1:]
    projects = [p.replace("проект ", "").strip() for p in projects]
    real_projects = await generate_response(prompts.GET_PROJECT_LIST_CHECK_PROJECT.format(text=request.text, projects=projects))
    real_projects_list = re.findall(r'Да:\s*(.*?)\s*(?:\n\n|$)', real_projects)
    return {"result": real_projects_list}

@app.post("/getprojectdetails")
async def getprojectdetails_route(request: TextRequest):
    if not request.projects:
        raise HTTPException(status_code=400, detail="Проекты не выбраны")
    
    final = {}
    for project in request.projects:
        prompts_list = [
            prompts.GET_PROJECT_DETAILS_AIM.format(text=request.text, project=project),
            prompts.GET_PROJECT_DETAILS_VALUE.format(text=request.text, project=project),
            prompts.GET_PROJECT_DETAILS_BUDGET.format(text=request.text, project=project),
            prompts.GET_PROJECT_DETAILS_ECO_EFFECT.format(text=request.text, project=project),
            prompts.GET_PROJECT_DETAILS_DEADLINE.format(text=request.text, project=project),
            prompts.GET_PROJECT_DETAILS_NEW_PLAN.format(text=request.text, project=project),
            prompts.GET_PROJECT_DETAILS_CONCLUSION.format(text=request.text, project=project),
        ]
        final[project] = "\n\n".join([await generate_response(p) for p in prompts_list])
        final[project] = final[project].replace("Конец ответа", "").replace('</s>', '').strip()
    
    return {"result": final}

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)