LongArticleTaskServer/applications/tasks/llm_tasks/candidate_account_process.py

import json
import traceback
from typing import List, Dict
from tqdm import tqdm

from applications.api import fetch_deepseek_completion
from applications.utils import ci_lower


class CandidateAccountProcessConst:

    INIT_STATUS = 0
    PROCESSING_STATUS = 1
    SUCCESS_STATUS = 2
    FAILED_STATUS = 99
    LACK_ARTICLE_STATUS = 11
    TITLE_TOO_LONG_STATUS = 14

    AVG_SCORE_THRESHOLD = 65

    ARTICLE_COUNT_THRESHOLD = 13
    AVG_TITLE_LENGTH_THRESHOLD = 45

    ACCOUNT_GOOD_STATUS = 1

    @staticmethod
    def generate_title_match_score_prompt(title_list):
        title_list_string = "\n".join(title_list)
        prompt = f"""
** 任务指令 **
    你是一名资深中文新闻编辑，需根据以下标准对一批标题进行主题匹配度评分（0-100分）

** 评估维度及权重 **
    1. 受众精准度（50%）
        正向匹配：存款/养老/健康/饮食/疾病警示/家庭伦理/近现代战争历史/老知青/奇闻异事
        负向排除：影视解说/文学解读/个人收藏（钱币/邮票）/机械科普/数码测评/电子游戏/时尚潮流/明星八卦/极限运动/学术研究/网络热梗/宠物饲养/音乐/棋牌

    2. 标题技法（40%）
        悬念设计：疑问句/省略号/反转结构（例："打开后瞬间愣住..."）
        情感强度：使用"痛心！""寒心！"等情绪词
        数据冲击：具体数字增强可信度（例："存款180万消失"）
        口语化表达：使用"涨知识了""别不当回事"等日常用语

    3. 内容调性（10%）
        煽情猎奇：家庭悲剧/离奇事件（例："棺材板挖出金条"）
        警示价值：健康建议/法律案例（例："三种食物禁止二次加热"）
        历史揭秘：人物秘闻/老照片故事
        爱国情怀：军事突破/资源发现（例："南极发现巨型粮仓"）

** 评分规则 **
    90-100分：同时满足3个维度且要素齐全，无负向内容
    70-89分：满足2个核心维度，无负向内容
    50-69分：仅满足受众群体正向匹配，无负向内容
    30-49分：存在轻微关联但要素缺失
    0-29分：完全无关或包含任意负向品类内容

** 待评估标题 **
    {title_list_string}

** 输出要求 **
    输出结果为JSON，仅输出这一批标题的评分，用数组 List 返回 [score1, score2, score3,...] 不要包含任何解释或说明。
"""
        return prompt


class CandidateAccountQualityScoreRecognizer(CandidateAccountProcessConst):
    def __init__(self, pool, log_client, trace_id):
        self.pool = pool
        self.log_client = log_client
        self.trace_id = trace_id

    async def get_task_list(self) -> List[Dict]:
        """
        get account tasks from the database
        """
        fetch_query = f"""
            select id, title_list, platform, account_id, account_name
            from crawler_candidate_account_pool
            where avg_score is null and status = {self.INIT_STATUS} and title_list is not null;
        """
        fetch_response = await self.pool.async_fetch(
            fetch_query,
        )
        return fetch_response

    async def update_account_status(
        self, account_id: int, ori_status: int, new_status: int
    ) -> int:
        """update account status"""
        update_query = f"""
                update crawler_candidate_account_pool
                set status = %s
                where id = %s and status = %s;
            """
        return await self.pool.async_save(
            update_query, (new_status, account_id, ori_status)
        )

    async def insert_account_into_crawler_queue(self, score_list: List[int], account: dict) -> None:
        """
        计算账号的得分置信区间下限，若置信区间下限的分数大于阈值，则认为是好的账号
        """
        if ci_lower(score_list) > self.AVG_SCORE_THRESHOLD:
            query = f"""
                insert into article_meta_accounts (platform, account_id, account_name, account_source, status)
                values (%s, %s, %s, %s, %s);
            """
            await self.pool.async_save(
                query=query,
                params=(account["platform"], account["account_id"], account["account_name"], 'ai_recognize', self.ACCOUNT_GOOD_STATUS)
            )

    async def score_for_each_account_by_llm(self, account):
        account_id = account["id"]
        # lock
        if not await self.update_account_status(
            account_id, self.INIT_STATUS, self.PROCESSING_STATUS
        ):
            return

        # start processing
        title_list = json.loads(account["title_list"])
        if (
            len(title_list) < self.ARTICLE_COUNT_THRESHOLD
            and account["platform"] == "toutiao"
        ):
            await self.update_account_status(
                account_id, self.PROCESSING_STATUS, self.LACK_ARTICLE_STATUS
            )
            return

        # 平均标题过长
        avg_title_length = sum([len(title) for title in title_list]) / len(title_list)
        if avg_title_length > self.AVG_TITLE_LENGTH_THRESHOLD:
            await self.update_account_status(
                account_id, self.PROCESSING_STATUS, self.TITLE_TOO_LONG_STATUS
            )
            return

        prompt = self.generate_title_match_score_prompt(title_list)
        try:
            completion = fetch_deepseek_completion(
                model="DeepSeek-V3", prompt=prompt, output_type="json"
            )
            avg_score = sum(completion) / len(completion)
            query = f"""
                update crawler_candidate_account_pool
                set score_list = %s, avg_score = %s, status = %s
                where id = %s and status = %s;
            """
            await self.pool.async_save(
                query=query,
                params=(
                    json.dumps(completion),
                    avg_score,
                    self.PROCESSING_STATUS,
                    account_id,
                    self.SUCCESS_STATUS,
                ),
            )
            # 判断置信区间下限, 并且插入账号
            await self.insert_account_into_crawler_queue(score_list=completion, account=account)

        except Exception as e:
            await self.log_client.log(
                contents={
                    "task": "candidate_account_analysis",
                    "trace_id": self.trace_id,
                    "function": "score_for_each_account_by_llm",
                    "message": "大模型识别账号失败",
                    "status": "fail",
                    "data": {
                        "error": str(e),
                        "title_list": json.dumps(title_list),
                    },
                }
            )
            await self.update_account_status(
                account_id, self.PROCESSING_STATUS, self.FAILED_STATUS
            )

    async def deal(self):
        task_list = await self.get_task_list()
        for task in tqdm(task_list, desc="use llm to analysis each account"):
            try:
                await self.score_for_each_account_by_llm(task)
            except Exception as e:
                await self.log_client.log(
                    contents={
                        "task": "candidate_account_analysis",
                        "trace_id": self.trace_id,
                        "function": "deal",
                        "status": "fail",
                        "data": {
                            "error": str(e),
                            "traceback": traceback.format_exc(),
                            "task": task,
                        },
                    }
                )