dataClean/app/summarizer.py

"""文章摘要生成器：对无摘要或短摘要文章调用 LLM 生成 AI 摘要"""
import logging
from datetime import datetime, timezone
from typing import List, Dict, Any

from sqlalchemy.orm import Session

from app.ai_client import ai_client
from app.rss_client import rss_client
from config import settings
from models import EnrichedArticle

logger = logging.getLogger(__name__)


SUMMARY_SYSTEM_PROMPT = """你是一位擅长阅读 RSS 新闻并提炼摘要的助手。
请用简洁流畅的中文总结文章核心内容，要求：
1. 长度控制在 {max_length} 个汉字以内。
2. 包含文章最重要的 1-3 个要点。
3. 不要添加个人评价，不要复述原文标题。
4. 若原文是英文，请用中文输出摘要。
"""


SUMMARY_USER_PROMPT_TEMPLATE = """请为以下文章生成摘要。

标题：{title}
作者：{author}
来源：{feed_title}

正文：
{content}
"""


def _needs_summary(article: EnrichedArticle) -> bool:
    """判断是否需要生成 AI 摘要"""
    if not article.ai_summary:
        return True
    original = article.original_summary or ""
    if len(original.strip()) < settings.MIN_ORIGINAL_SUMMARY_LENGTH:
        return True
    return False


def _prepare_content(raw_content: str, max_chars: int = 8000) -> str:
    """清洗并截断正文，避免超过 LLM 上下文"""
    text = raw_content or ""
    # 简单去除多余空白
    text = " ".join(text.split())
    return text[:max_chars]


def _generate_summary(article: EnrichedArticle) -> str:
    """调用 LLM 生成单篇文章摘要"""
    content = _prepare_content(article.content or article.original_summary or "")
    if not content.strip():
        # 如果连原始摘要都没有，只能基于标题生成
        content = article.title or ""

    system_prompt = SUMMARY_SYSTEM_PROMPT.format(max_length=settings.MAX_AI_SUMMARY_LENGTH)
    user_prompt = SUMMARY_USER_PROMPT_TEMPLATE.format(
        title=article.title or "",
        author=article.author or "",
        feed_title=article.feed_title or "",
        content=content,
    )

    try:
        summary = ai_client.chat_completion(
            system_prompt=system_prompt,
            user_prompt=user_prompt,
            temperature=0.3,
        )
        return summary[: settings.MAX_AI_SUMMARY_LENGTH]
    except Exception as exc:
        logger.error("生成 article_id=%d 摘要失败: %s", article.rk_article_id, exc)
        return ""


def _article_from_rss(raw: Dict[str, Any]) -> Dict[str, Any]:
    """把 rssKeeper 返回的文章转换为可写入 enriched 表的字典"""
    published_at = raw.get("published_at")
    if isinstance(published_at, str):
        try:
            published_at = datetime.fromisoformat(published_at.replace("Z", "+00:00"))
        except Exception:
            published_at = None

    return {
        "rk_article_id": raw["id"],
        "title": raw.get("title", "") or "",
        "link": raw.get("link", "") or "",
        "feed_id": raw.get("feed_id", 0),
        "feed_title": raw.get("feed_title", "") or "",
        "feed_category": raw.get("category", "") or "",
        "author": raw.get("author", "") or "",
        "published_at": published_at,
        "original_summary": raw.get("summary", "") or "",
        "content": raw.get("content", "") or "",
    }


def fetch_and_summarize(db: Session, hours: int = 24, limit: int = 200) -> Dict[str, int]:
    """
    拉取最近文章，补充 AI 摘要。
    返回统计信息：{"fetched": x, "created": y, "summarized": z}
    """
    articles = rss_client.fetch_recent(hours=hours, limit=limit)
    if not articles:
        logger.info("未拉取到新文章")
        return {"fetched": 0, "created": 0, "summarized": 0}

    stats = {"fetched": len(articles), "created": 0, "summarized": 0}

    for raw in articles:
        data = _article_from_rss(raw)
        article = db.query(EnrichedArticle).filter(
            EnrichedArticle.rk_article_id == data["rk_article_id"]
        ).first()

        if article is None:
            article = EnrichedArticle(**data)
            db.add(article)
            db.flush()
            stats["created"] += 1
        else:
            # 更新已有记录的基础字段
            article.title = data["title"] or article.title
            article.link = data["link"] or article.link
            article.feed_title = data["feed_title"] or article.feed_title
            article.feed_category = data["feed_category"] or article.feed_category
            article.author = data["author"] or article.author
            article.published_at = data["published_at"] or article.published_at
            article.original_summary = data["original_summary"] or article.original_summary
            article.content = data["content"] or article.content
            article.fetched_at = datetime.now(timezone.utc)

        if _needs_summary(article):
            ai_summary = _generate_summary(article)
            if ai_summary:
                article.ai_summary = ai_summary
                stats["summarized"] += 1

        # 每 10 篇提交一次，避免长时间事务
        if stats["summarized"] % 10 == 0:
            db.commit()

    db.commit()
    logger.info(
        "摘要任务完成: fetched=%d, created=%d, summarized=%d",
        stats["fetched"], stats["created"], stats["summarized"]
    )
    return stats