mirror of
https://github.com/stardrophere/InsightRadar.git
synced 2026-06-06 01:57:51 +08:00
login+ai cluster
This commit is contained in:
stardrophere
@@ -1,62 +1,222 @@
|
||||
# app/services/fetcher_service.py
|
||||
import os
|
||||
import hashlib
|
||||
import httpx
|
||||
from dotenv import load_dotenv
|
||||
from datetime import timedelta
|
||||
|
||||
import httpx
|
||||
import json
|
||||
import numpy as np
|
||||
from dotenv import load_dotenv
|
||||
from sklearn.metrics.pairwise import cosine_similarity
|
||||
from sentence_transformers import SentenceTransformer
|
||||
from app.database import SessionLocal
|
||||
from app.models.models import (
|
||||
InfoSource, TrendingEvent, NewsArticle, DataSyncTask, TaskStatus,
|
||||
HeadlineRevision, RankingLog, SourceType, utcnow
|
||||
HeadlineRevision, RankingLog, SourceType, utcnow, UnifiedEvent
|
||||
)
|
||||
|
||||
# ==========================================
|
||||
# 环境变量与全局配置
|
||||
# ==========================================
|
||||
# 加载环境变量
|
||||
load_dotenv()
|
||||
|
||||
# 从环境变量获取 API 基础地址，如果没有配置则提供默认回退地址
|
||||
hf_token = os.getenv("HF_TOKEN")
|
||||
SIMILARITY_THRESHOLD = float(os.getenv("SIMILARITY_THRESHOLD", 0.72))
|
||||
API_BASE_URL = os.getenv("API_BASE_URL", "https://newsnow.busiyi.world/api/s")
|
||||
EMBEDDING_MODEL_PATH = os.getenv("EMBEDDING_MODEL_PATH", "")
|
||||
|
||||
print("正在加载 BAAI/bge-m3 向量模型...")
|
||||
# 全局单例
|
||||
embedder_model = SentenceTransformer(EMBEDDING_MODEL_PATH, local_files_only=True, device="cuda")
|
||||
print("模型加载完成。")
|
||||
|
||||
|
||||
def generate_md5(text: str) -> str:
|
||||
"""
|
||||
生成32位MD5哈希值
|
||||
维护说明：
|
||||
各个平台（微博、知乎、微信等）返回的原始 ID 格式千奇百怪（有长数字、有UUID、有URL）。
|
||||
为了方便数据库建立统一的高性能唯一索引（UniqueConstraint），
|
||||
我们统一将其转为长度固定的 32 位 MD5 字符串作为 external_id。
|
||||
"""
|
||||
"""生成32位MD5哈希值作为全局唯一指纹"""
|
||||
return hashlib.md5(text.encode('utf-8')).hexdigest()
|
||||
|
||||
|
||||
def generate_embedding_json(text: str) -> str:
|
||||
"""辅助函数：调用大模型生成向量，并序列化为 JSON 字符串"""
|
||||
raw_vec = embedder_model.encode([text], normalize_embeddings=True)[0]
|
||||
truncated_vec = [round(float(x), 5) for x in raw_vec]
|
||||
return json.dumps(truncated_vec, separators=(',', ':'))
|
||||
|
||||
|
||||
def match_or_create_unified_event(db, title: str, embedding_json: str) -> int:
|
||||
"""
|
||||
辅助函数：大事件聚类中枢。
|
||||
拿着新计算的向量去数据库里碰，碰到了就返回老 ID，碰不到就建新的。
|
||||
"""
|
||||
# 提取刚算出来的向量
|
||||
new_vec = np.array(json.loads(embedding_json))
|
||||
|
||||
# 只取最近 3 天的活跃大事件进行比对
|
||||
three_days_ago = utcnow() - timedelta(days=3)
|
||||
recent_events = db.query(UnifiedEvent).filter(
|
||||
UnifiedEvent.created_at >= three_days_ago
|
||||
).order_by(UnifiedEvent.created_at.desc()).limit(200).all()
|
||||
|
||||
if recent_events:
|
||||
valid_events = [ev for ev in recent_events if ev.center_embedding]
|
||||
if valid_events:
|
||||
event_vectors = [json.loads(ev.center_embedding) for ev in valid_events]
|
||||
|
||||
# 批量矩阵计算相似度
|
||||
sim_scores = cosine_similarity([new_vec], event_vectors)[0]
|
||||
max_idx = np.argmax(sim_scores)
|
||||
|
||||
if sim_scores[max_idx] >= SIMILARITY_THRESHOLD:
|
||||
matched_event = valid_events[max_idx]
|
||||
|
||||
matched_event.hot_score += 1
|
||||
return matched_event.id
|
||||
|
||||
# 没匹配到，创建一个新的统一大事件
|
||||
new_unified = UnifiedEvent(
|
||||
unified_title=title,
|
||||
center_embedding=embedding_json,
|
||||
hot_score=1 # 初始热度
|
||||
)
|
||||
db.add(new_unified)
|
||||
db.flush() # 获取自增的主键 ID
|
||||
return new_unified.id
|
||||
|
||||
|
||||
def process_hot_trend_item(db, source, item, index: int, external_id: str):
|
||||
"""
|
||||
处理【热搜/短新闻】的业务逻辑，现已加入 AI 聚类功能
|
||||
"""
|
||||
title = item.get("title")
|
||||
item_url = item.get("url", "")
|
||||
|
||||
existing_event = db.query(TrendingEvent).filter(
|
||||
TrendingEvent.source_id == source.id,
|
||||
TrendingEvent.external_id == external_id
|
||||
).first()
|
||||
|
||||
event_to_log = None
|
||||
|
||||
# 核心逻辑：查重后再决定是否调用模型
|
||||
if existing_event:
|
||||
# 场景 A1：老熟人
|
||||
if existing_event.current_headline != title:
|
||||
# 标题被暗改，此时需要重新算一次 Embedding
|
||||
new_embedding_json = generate_embedding_json(title)
|
||||
|
||||
revision = HeadlineRevision(
|
||||
event_id=existing_event.id,
|
||||
previous_headline=existing_event.current_headline,
|
||||
revised_headline=title
|
||||
)
|
||||
db.add(revision)
|
||||
existing_event.current_headline = title
|
||||
existing_event.title_embedding = new_embedding_json # 更新为新标题的语义向量
|
||||
# 注：这里不改变它所属的 unified_event_id，因为大体还是同一件事
|
||||
|
||||
existing_event.current_ranking = index
|
||||
existing_event.event_url = item_url
|
||||
event_to_log = existing_event
|
||||
|
||||
else:
|
||||
# 场景 A2：这是一条彻底的全新热搜
|
||||
# 1. 计算向量
|
||||
new_embedding_json = generate_embedding_json(title)
|
||||
|
||||
# 2. 扔进聚类中枢找归宿
|
||||
matched_event_id = match_or_create_unified_event(db, title, new_embedding_json)
|
||||
|
||||
# 3. 落库
|
||||
new_event = TrendingEvent(
|
||||
source_id=source.id,
|
||||
external_id=external_id,
|
||||
current_headline=title,
|
||||
event_url=item_url,
|
||||
current_ranking=index,
|
||||
title_embedding=new_embedding_json, # 存入向量
|
||||
unified_event_id=matched_event_id # 挂载到大事件下
|
||||
)
|
||||
db.add(new_event)
|
||||
db.flush()
|
||||
event_to_log = new_event
|
||||
|
||||
# 强制记录排名轨迹
|
||||
rank_log = RankingLog(
|
||||
event_id=event_to_log.id,
|
||||
ranking_position=index
|
||||
)
|
||||
db.add(rank_log)
|
||||
|
||||
|
||||
def process_rss_feed_item(db, source, item, external_id: str):
|
||||
"""
|
||||
处理【长文章/传统订阅】分支的核心业务逻辑 (写入 NewsArticle 表)
|
||||
"""
|
||||
title = item.get("title")
|
||||
item_url = item.get("url", "")
|
||||
|
||||
existing_article = db.query(NewsArticle).filter(
|
||||
NewsArticle.source_id == source.id,
|
||||
NewsArticle.external_id == external_id
|
||||
).first()
|
||||
|
||||
if existing_article:
|
||||
# 文章若存在，仅更新基础字段
|
||||
existing_article.article_title = title
|
||||
existing_article.article_url = item_url
|
||||
else:
|
||||
# 全新文章入库
|
||||
new_article = NewsArticle(
|
||||
source_id=source.id,
|
||||
external_id=external_id,
|
||||
article_title=title,
|
||||
article_url=item_url,
|
||||
)
|
||||
db.add(new_article)
|
||||
|
||||
|
||||
def process_source_data(db, source, items: list) -> int:
|
||||
"""
|
||||
数据清洗与路由分发层：
|
||||
遍历 API 返回的 items，生成唯一指纹，并路由到不同的处理模块。
|
||||
返回成功处理的条目数量。
|
||||
"""
|
||||
saved_count = 0
|
||||
platform_id = source.home_url
|
||||
|
||||
for index, item in enumerate(items, 1):
|
||||
title = item.get("title")
|
||||
if not title:
|
||||
continue
|
||||
|
||||
item_url = item.get("url", "")
|
||||
|
||||
# ID 兜底策略：接口ID -> URL -> Title
|
||||
raw_id = item.get("id") or item_url or title
|
||||
external_id = generate_md5(f"{platform_id}_{raw_id}")
|
||||
|
||||
# 核心路由分流
|
||||
if source.source_type in (SourceType.HOT_TREND, SourceType.API):
|
||||
process_hot_trend_item(db, source, item, index, external_id)
|
||||
elif source.source_type == SourceType.RSS_FEED:
|
||||
process_rss_feed_item(db, source, item, external_id)
|
||||
|
||||
saved_count += 1
|
||||
|
||||
return saved_count
|
||||
|
||||
|
||||
async def fetch_and_save_trending_data():
|
||||
"""
|
||||
核心定时任务：从数据库读取信息源 -> 抓取API -> 解析 -> 根据业务类型分流存入对应的数据库表
|
||||
|
||||
执行流程：
|
||||
1. 查询所有配置为“已启用”的信息源 (is_enabled == True)。
|
||||
2. 伪装 HTTP 请求头，规避目标服务器的反爬机制。
|
||||
3. 遍历解析数据，生成 MD5 唯一指纹进行全局去重。
|
||||
4. 核心路由分流：
|
||||
- 若源为 HOT_TREND/API，按热搜逻辑处理（记录名次轨迹、标题变更）。
|
||||
- 若源为 RSS_FEED，按长文章逻辑处理（忽略名次，直接落库）。
|
||||
5. 严格的事务管理：成功则统一提交，报错则回滚业务数据并独立提交错误日志。
|
||||
调度层：负责网络请求、数据库事务管理和异常监控隔离。
|
||||
"""
|
||||
print(f"[{utcnow()}] 开始执行定时抓取任务...")
|
||||
|
||||
# 使用上下文管理器确保数据库连接池正确获取和归还连接
|
||||
with SessionLocal() as db:
|
||||
# 1. 动态获取抓取源。
|
||||
# 优势：在后台修改数据库的信息源开关，下一次定时任务立刻生效，无需重启服务。
|
||||
# 获取启用的信息源
|
||||
sources = db.query(InfoSource).filter(InfoSource.is_enabled == True).all()
|
||||
if not sources:
|
||||
print("没有找到启用的信息源，任务结束。")
|
||||
return
|
||||
|
||||
# 2. 伪装成真实的浏览器 HTTP 请求头
|
||||
# 维护注意：如果抓取接口返回 403 Forbidden，通常是这里的反爬策略失效了，需要更新 User-Agent 或 Cookie
|
||||
# 伪装请求头，规避反爬
|
||||
custom_headers = {
|
||||
"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/145.0.0.0 Safari/537.36",
|
||||
"Accept": "application/json, text/plain, */*",
|
||||
@@ -64,157 +224,40 @@ async def fetch_and_save_trending_data():
|
||||
"Origin": "https://newsnow.busiyi.world"
|
||||
}
|
||||
|
||||
# 复用异步 HTTP 客户端，比每次循环新建 Client 性能更高
|
||||
async with httpx.AsyncClient(timeout=15.0, headers=custom_headers) as client:
|
||||
for source in sources:
|
||||
# platform_id 对应第三方接口的入参标识，如 "weibo", "zhihu" 等
|
||||
platform_id = source.home_url
|
||||
if not platform_id:
|
||||
continue
|
||||
|
||||
# ==========================================
|
||||
# 【技术债预警 / TODO】
|
||||
# 目前无论 source_type 是什么，都统一请求了这个 JSON API。
|
||||
# 未来如果加入了真正的外部 RSS 订阅源（返回的是 XML 格式），
|
||||
# 这里需要增加判断逻辑：如果是 RSS_FEED，应当使用 feedparser 库去解析 XML，而不是用 httpx 获取 JSON。
|
||||
# ==========================================
|
||||
url = f"{API_BASE_URL}?id={platform_id}&latest"
|
||||
|
||||
# 初始化本次特定信息源抓取任务的系统监控日志
|
||||
# 初始化监控日志
|
||||
task_log = DataSyncTask(source_id=source.id, items_fetched=0)
|
||||
|
||||
try:
|
||||
# 发起请求并校验 HTTP 状态码 (非 2xx 会抛出异常进入 except 块)
|
||||
# 发起网络请求
|
||||
response = await client.get(url)
|
||||
response.raise_for_status()
|
||||
data_json = response.json()
|
||||
|
||||
items = data_json.get("items", [])
|
||||
saved_count = 0
|
||||
|
||||
for index, item in enumerate(items, 1):
|
||||
title = item.get("title")
|
||||
if not title:
|
||||
continue
|
||||
# 调用数据处理层
|
||||
saved_count = process_source_data(db, source, items)
|
||||
|
||||
item_url = item.get("url", "")
|
||||
|
||||
# 3. ID 兜底与去重策略
|
||||
# 优先用接口自带的 ID -> 没有则用 URL 代替 -> 最差情况用 title 兜底
|
||||
raw_id = item.get("id") or item_url or title
|
||||
# 组合“平台标识+原始ID”算出全局唯一的 MD5 外部标识
|
||||
external_id = generate_md5(f"{platform_id}_{raw_id}")
|
||||
|
||||
# ==========================================
|
||||
# 4. 核心数据分流路由
|
||||
# 根据信息源的业务类型，将数据推入不同的物理表
|
||||
# ==========================================
|
||||
|
||||
if source.source_type in (SourceType.HOT_TREND, SourceType.API):
|
||||
# --------------------------------------------------
|
||||
# 分支 A：热搜/短新闻逻辑 -> 写入 TrendingEvent 表
|
||||
# --------------------------------------------------
|
||||
existing_event = db.query(TrendingEvent).filter(
|
||||
TrendingEvent.source_id == source.id,
|
||||
TrendingEvent.external_id == external_id
|
||||
).first()
|
||||
|
||||
event_to_log = None # 临时指针，用于后续绑定排名轨迹
|
||||
|
||||
if existing_event:
|
||||
# 场景 A1：该热搜已经在数据库中
|
||||
|
||||
# 监控并记录“标题暗改”（常见于热搜公关介入）
|
||||
if existing_event.current_headline != title:
|
||||
revision = HeadlineRevision(
|
||||
event_id=existing_event.id,
|
||||
previous_headline=existing_event.current_headline,
|
||||
revised_headline=title
|
||||
)
|
||||
db.add(revision)
|
||||
existing_event.current_headline = title # 覆盖为主表最新标题
|
||||
|
||||
# 更新当前的实时排名和 URL
|
||||
existing_event.current_ranking = index
|
||||
existing_event.event_url = item_url
|
||||
event_to_log = existing_event
|
||||
else:
|
||||
# 场景 A2：发现全新热搜
|
||||
new_event = TrendingEvent(
|
||||
source_id=source.id,
|
||||
external_id=external_id,
|
||||
current_headline=title,
|
||||
event_url=item_url,
|
||||
current_ranking=index,
|
||||
)
|
||||
db.add(new_event)
|
||||
# db.flush() 是关键：它将数据推给数据库生成了自增的主键 ID，但尚未最终 commit。
|
||||
# 拿到合法的 event_to_log.id
|
||||
db.flush()
|
||||
event_to_log = new_event
|
||||
|
||||
# 排名轨迹强制记录
|
||||
# 只要抓到了热搜（无论新旧），必须打点记录当前名次，用于前端绘制排名趋势图
|
||||
rank_log = RankingLog(
|
||||
event_id=event_to_log.id,
|
||||
ranking_position=index
|
||||
)
|
||||
db.add(rank_log)
|
||||
|
||||
elif source.source_type == SourceType.RSS_FEED:
|
||||
# --------------------------------------------------
|
||||
# 分支 B：长文章/传统订阅逻辑 -> 写入 NewsArticle 表
|
||||
# --------------------------------------------------
|
||||
existing_article = db.query(NewsArticle).filter(
|
||||
NewsArticle.source_id == source.id,
|
||||
NewsArticle.external_id == external_id
|
||||
).first()
|
||||
|
||||
if existing_article:
|
||||
# 文章如果已存在，通常只需要更新基础字段（文章一般不涉及排名起伏）
|
||||
existing_article.article_title = title
|
||||
existing_article.article_url = item_url
|
||||
# 预留位置：如果以后接口返回了摘要，可以在这里 update existing_article.original_summary
|
||||
else:
|
||||
# 全新文章入库
|
||||
new_article = NewsArticle(
|
||||
source_id=source.id,
|
||||
external_id=external_id,
|
||||
article_title=title,
|
||||
article_url=item_url,
|
||||
# original_summary=item.get("desc", ""),
|
||||
# author_name=item.get("author", "")
|
||||
)
|
||||
db.add(new_article)
|
||||
|
||||
|
||||
|
||||
saved_count += 1
|
||||
|
||||
# --------------------------------------------------
|
||||
# 5. 业务事务成功提交
|
||||
# --------------------------------------------------
|
||||
# 只有当前平台（source）的所有 item 都顺畅走完，才标记成功
|
||||
# 业务事务成功提交
|
||||
task_log.items_fetched = saved_count
|
||||
task_log.task_status = TaskStatus.SUCCESS
|
||||
db.add(task_log)
|
||||
|
||||
# 统一将当前信息源爬取的所有业务数据持久化到硬盘
|
||||
db.commit()
|
||||
print(f"[{source.source_name}] ({source.source_type}) 成功抓取并更新了 {saved_count} 条数据")
|
||||
|
||||
except Exception as e:
|
||||
# --------------------------------------------------
|
||||
# 6. 异常拦截与错误隔离机制
|
||||
# --------------------------------------------------
|
||||
# 回滚本次抓取的全部脏数据，
|
||||
# 异常拦截与错误隔离
|
||||
db.rollback()
|
||||
|
||||
# 错误日志记下来
|
||||
task_log.task_status = TaskStatus.ERROR
|
||||
task_log.error_trace = str(e)
|
||||
db.add(task_log)
|
||||
|
||||
|
||||
db.commit()
|
||||
print(f"[{source.source_name}] 抓取失败: {e}")
|
||||
print(f"[{source.source_name}] 抓取失败: {e}")
|
||||
|
||||
@@ -0,0 +1,104 @@
|
||||
# app/services/summary_service.py
|
||||
import os
|
||||
import json
|
||||
from datetime import timedelta
|
||||
from openai import AsyncOpenAI
|
||||
|
||||
from app.database import SessionLocal
|
||||
from app.models.models import UnifiedEvent, TrendingEvent, InfoSource, utcnow
|
||||
from app.prompts.summary_prompts import (
|
||||
SUMMARY_SYSTEM_PROMPT,
|
||||
SUMMARY_USER_PROMPT_TEMPLATE,
|
||||
)
|
||||
|
||||
HOT_SCORE_THRESHOLD = int(os.getenv("HOT_SCORE_THRESHOLD", 3))
|
||||
AI_API_KEY = os.getenv("AI_API_KEY", '')
|
||||
|
||||
# 1. 初始化异步客户端 (全局复用)
|
||||
deepseek_client = AsyncOpenAI(
|
||||
api_key=AI_API_KEY,
|
||||
base_url="https://api.deepseek.com"
|
||||
)
|
||||
|
||||
|
||||
async def call_llm_for_summary(platform_data_text: str) -> dict:
|
||||
"""调用 DeepSeek 生成统一标题和多平台视角摘要"""
|
||||
prompt = SUMMARY_USER_PROMPT_TEMPLATE.format(
|
||||
platform_data_text=platform_data_text
|
||||
)
|
||||
|
||||
# await
|
||||
response = await deepseek_client.chat.completions.create(
|
||||
model="deepseek-chat",
|
||||
messages=[
|
||||
{"role": "system", "content": SUMMARY_SYSTEM_PROMPT},
|
||||
{"role": "user", "content": prompt}
|
||||
],
|
||||
response_format={"type": "json_object"},
|
||||
temperature=1
|
||||
)
|
||||
|
||||
result_text = response.choices[0].message.content
|
||||
return json.loads(result_text)
|
||||
|
||||
|
||||
async def generate_unified_summaries():
|
||||
"""定时任务：扫描高热度事件并生成/更新摘要"""
|
||||
print(f"[{utcnow()}] 开始执行 DeepSeek 摘要生成任务...")
|
||||
|
||||
with SessionLocal() as db:
|
||||
recent_threshold = utcnow() - timedelta(days=3)
|
||||
|
||||
# 必须满足：热度达标 AND (当前热度 > 上次生成摘要时的热度) AND 近期活跃
|
||||
events = db.query(UnifiedEvent).filter(
|
||||
UnifiedEvent.hot_score >= HOT_SCORE_THRESHOLD,
|
||||
UnifiedEvent.hot_score > UnifiedEvent.last_summarized_trends_count,
|
||||
UnifiedEvent.created_at >= recent_threshold
|
||||
).all()
|
||||
|
||||
if not events:
|
||||
print("当前没有需要更新摘要的大事件，任务结束。")
|
||||
return
|
||||
|
||||
for event in events:
|
||||
# 联合查询获取该事件在各平台的子新闻
|
||||
trends = db.query(TrendingEvent, InfoSource.source_name) \
|
||||
.join(InfoSource, TrendingEvent.source_id == InfoSource.id) \
|
||||
.filter(TrendingEvent.unified_event_id == event.id) \
|
||||
.all()
|
||||
|
||||
if not trends:
|
||||
continue
|
||||
|
||||
# 按平台归类标题并去重
|
||||
platform_dict = {}
|
||||
for trend_record, source_name in trends:
|
||||
if source_name not in platform_dict:
|
||||
platform_dict[source_name] = set()
|
||||
platform_dict[source_name].add(trend_record.current_headline)
|
||||
|
||||
# 组装给大模型的 Prompt 数据
|
||||
prompt_lines = [f"【{platform}】: {', '.join(headlines)}" for platform, headlines in platform_dict.items()]
|
||||
platform_data_text = "\n".join(prompt_lines)
|
||||
|
||||
try:
|
||||
# 调用封装好的异步函数
|
||||
llm_result = await call_llm_for_summary(platform_data_text)
|
||||
|
||||
if "unified_title" in llm_result:
|
||||
event.unified_title = llm_result["unified_title"]
|
||||
if "ai_comprehensive_summary" in llm_result:
|
||||
event.ai_comprehensive_summary = llm_result["ai_comprehensive_summary"]
|
||||
|
||||
# 成功后更新水位线
|
||||
# 将最后一次总结时的热搜数量，更新为当前最新的 hot_score
|
||||
event.last_summarized_trends_count = event.hot_score
|
||||
|
||||
print(f"成功更新大事件 ID {event.id} 的深度摘要 (当前热度: {event.hot_score})。")
|
||||
|
||||
except Exception as e:
|
||||
print(f"大事件 ID {event.id} 摘要生成失败: {e}")
|
||||
continue
|
||||
|
||||
# 提交事务
|
||||
db.commit()
|
||||
Reference in New Issue
Block a user