feat(llm): classify 前置 + 黑名单 drop 删文章 + 排版用 .diary-para

- enrichment._enrich_classify 前置,返回 (drop, categories)
  - 注入 {blocklist} 占位符到 prompt(全局 + per-source 合并)
  - drop=True → 整篇 DELETE,后 3 步直接 skip
  - 兜底:即使 LLM 没正确返回 drop 字段,本地也匹配一次
- enrichment._enrich_format 排版段落 class 名固定为 diary-para
  - CSS 仍内联到 style,前端 .diary-para 兜底
- enrichment._merge_blocklist: 全局 + per-source 合并去重保序
- schemas/llm.LlmSettingOut/Update 暴露 blocklist_tags
- DEFAULT_PROMPTS.classify_prompt 加 {blocklist} + drop 字段说明

backend/app/schemas/llm.py

@@ -18,6 +18,8 @@ class LlmSettingOut(BaseModel):
     image_model: str = "agnes-image-2.1-flash"
     interval_sec: float = 2.0
     enabled: bool = True
+    # 全局屏蔽分类标签;与 sources.blocklist_tags 合并后注入 classify prompt
+    blocklist_tags: list[str] = []
     updated_at: datetime | None = None
 
 
@@ -33,6 +35,7 @@ class LlmSettingUpdate(BaseModel):
     image_model: str | None = Field(default=None, min_length=1, max_length=64)
     interval_sec: float | None = Field(default=None, ge=0.0, le=60.0)
     enabled: bool | None = None
+    blocklist_tags: list[str] | None = None
 
 
 # === 默认提示词(模板,用户可改)===
@@ -50,7 +53,10 @@ DEFAULT_PROMPTS = {
     "classify_prompt": (
         "你是新闻分类助手。请阅读以下新闻,返回 2-5 个最相关的分类标签(多标签)。\n"
         "可选标签(可自由组合,不限于此): 时政 / 经济 / 科技 / 军事 / 社会 / 国际 / 体育 / 文化 / 环境 / 健康 / 金融 / 能源 / 气候\n"
-        "严格要求:只返回 JSON,形如 {\"categories\": [\"时政\", \"国际\", \"经济\"]},不要其他内容。\n\n"
+        "黑名单分类(若新闻属于或主要围绕这些领域,务必将 drop 设为 true): {blocklist}\n"
+        "严格要求:只返回 JSON,形如 {\"categories\": [\"时政\", \"国际\", \"经济\"], \"drop\": false},"
+        "若新闻属于或主要围绕黑名单中的任何分类,将 drop 设为 true 并把该分类放入 categories。"
+        "不要其他内容。\n\n"
         "标题:{title}\n摘要:{summary}\n正文(节选):{body}\n"
     ),
     "commentary_prompt": (

backend/app/services/llm/enrichment.py

@@ -1,8 +1,8 @@
 """LLM 智能增强服务(翻译后调)。
 
-4 个独立任务:
+4 个独立任务(按顺序):
-  1. format   — 排版译文(写入 body_zh_formatted)
+  1. classify — 分类 + 黑名单 gate(命中则删文章,后 3 步跳过)
-  2. classify — 分类(写入 category,多标签)
+  2. format   — 排版译文(写入 body_zh_formatted,容器用 .article-body + 段落 .diary-para)
   3. image    — 生成插图(写入 image_ai_url,prompt 用正文第一段)
   4. commentary — 写点评(写入 commentary)
 
@@ -11,10 +11,15 @@
 - 字号: 17px
 - 行高: 1.7
 - 颜色: #3e3e3e
-- 段落: margin-bottom 1.5em(自动空一行)
+- 段落: margin-bottom 1.5em(自动空一行);class 名固定为 diary-para
+
+黑名单机制:
+- classify 任务合并 llm_settings.blocklist_tags(全局) + source.blocklist_tags(per-source)
+- 注入到 prompt 的 {blocklist} 占位符,LLM 返回 {"drop": true, "categories": [...]} 则删文章
+- 合并去重后为空 → classify 任务只产出 categories 不产出 drop
 
 设计:
-- 任务入口: enrich_article(article_id, settings_row)
+- 任务入口: enrich_article(article_id)
 - 任务间互不影响:每个任务独立 try/except + 写 status
 - 全部任务共走 LlmClient 的全局限速
 - 若设置 enabled=False,只跳过(不调 LLM)
@@ -31,6 +36,7 @@ from sqlalchemy import select
 from app.database import AsyncSessionLocal
 from app.models.article import Article
 from app.models.llm_setting import LlmSetting
+from app.models.source import Source
 from app.schemas.llm import get_default_prompts
 from app.services.llm.client import LlmClient
 
@@ -48,6 +54,9 @@ ARTICLE_BODY_LINE_HEIGHT = "1.7"
 ARTICLE_BODY_COLOR = "#3e3e3e"
 ARTICLE_BODY_P_MARGIN_BOTTOM = "1.5em"
 
+# === 排版段落 class 名(项目级固定,前端 .diary-para 兜底)===
+DIARY_PARA_CLASS = "diary-para"
+
 # === 插图默认尺寸(适中,不再用 1024x768)===
 # 写死到 enrichment 里,行为稳定;setting.image_size 仍可由用户在 UI 改,
 # 但默认行为不依赖它,避免意外被改成很大。
@@ -107,8 +116,13 @@ async def _enrich_format(article: Article, setting: LlmSetting, client: LlmClien
         temperature=0.3,
         max_tokens=2000,
     )
-    # 极简 HTML 包裹:按段切 + <p>,整体包到带固定 CSS 的 <div> 里
+    # 段落 class 名固定为 diary-para(项目级固定,前端 .diary-para 兜底);
-    parts = [f"<p>{p.strip()}</p>" for p in text.split("\n\n") if p.strip()]
+    # CSS 仍内联到 style,保证分享/导出场景不丢。
+    parts = [
+        f'<p class="{DIARY_PARA_CLASS}">{p.strip()}</p>'
+        for p in text.split("\n\n")
+        if p.strip()
+    ]
     if not parts:
         article.body_zh_formatted = None
     else:
@@ -129,30 +143,67 @@ def _wrap_article_body(inner_html: str) -> str:
     )
     # 段落样式也内联,保证 v-html 渲染时一定生效
     p_style = f"margin:0 0 {ARTICLE_BODY_P_MARGIN_BOTTOM} 0;"
-    inner_with_p_style = inner_html.replace("<p>", f'<p style="{p_style}">')
+    # 内层 HTML 的 <p class="diary-para" ...> 形式;把 style 插到 class 后面
+    inner_with_p_style = inner_html.replace(
+        f'<p class="{DIARY_PARA_CLASS}">',
+        f'<p class="{DIARY_PARA_CLASS}" style="{p_style}">',
+    )
     return f'<div class="article-body" style="{inline_style}">{inner_with_p_style}</div>'
 
 
-# === 单任务:classify ===
+# === 单任务:classify (含黑名单 drop gate) ===
-async def _enrich_classify(article: Article, setting: LlmSetting, client: LlmClient) -> None:
+async def _enrich_classify(
+    article: Article,
+    setting: LlmSetting,
+    client: LlmClient,
+    blocklist: list[str],
+) -> tuple[bool, list[str]]:
+    """分类 + 黑名单判断。
+
+    返回 (drop, categories):
+      - drop=True  → 整篇文章应删除(分类命中 blocklist)
+      - categories → 写入 article.category 的多标签列表
+    """
     template = setting.classify_prompt or get_default_prompts()["classify_prompt"]
-    # 老 prompt 可能只支持 {title}/{summary},不支持 {body} —— _safe_format 兜底
+    # 老 prompt 可能只支持 {title}/{summary},不支持 {body} / {blocklist} —— _safe_format 兜底
     vars_ = {
         "title": (article.title_zh or article.title)[:200],
         "summary": (article.summary_zh or "")[:400],
         "body": (article.body_zh_text or "")[:1500],
+        "blocklist": "、".join(blocklist) if blocklist else "(无)",
     }
     prompt = _safe_format(template, vars_)
     result = await client.classify_json(
         system="你是新闻分类助手,只返回 JSON。",
         user=prompt,
     )
-    cats = result.get("categories") or result.get("tags") or []
+    cats_raw = result.get("categories") or result.get("tags") or []
-    if isinstance(cats, list) and cats:
+    cats: list[str] = []
-        # 多标签(2-5 个),逗号分隔存到 category 字段(已有索引)
+    if isinstance(cats_raw, list):
-        joined = ",".join(str(c).strip() for c in cats[:DEFAULT_IMAGE_MAX_TAGS] if str(c).strip())
+        cats = [str(c).strip() for c in cats_raw[:DEFAULT_IMAGE_MAX_TAGS] if str(c).strip()]
-        article.category = joined[:64] or None
+    # 兼容两种来源:LLM 自己判断的 drop 字段,或后端兜底检查命中
-    article.classify_status = "ok"
+    drop_flag = bool(result.get("drop"))
+    if not drop_flag and blocklist:
+        # 兜底:即使 LLM 没正确返回 drop 字段,我们也用本地匹配兜底
+        bl_set = {b.strip() for b in blocklist if b and b.strip()}
+        drop_flag = any(c in bl_set for c in cats)
+    return drop_flag, cats
+
+
+def _merge_blocklist(setting: LlmSetting, source: Source | None) -> list[str]:
+    """合并全局 + per-source blocklist,去重保序。"""
+    out: list[str] = []
+    seen: set[str] = set()
+    sources: list[list[str]] = [setting.blocklist_tags or []]
+    if source is not None:
+        sources.append(source.blocklist_tags or [])
+    for src in sources:
+        for t in src:
+            t = (t or "").strip()
+            if t and t not in seen:
+                seen.add(t)
+                out.append(t)
+    return out
 
 
 # === 单任务:image ===
@@ -205,6 +256,10 @@ async def _enrich_commentary(article: Article, setting: LlmSetting, client: LlmC
 async def enrich_article(article_id: int) -> dict[str, str]:
     """对单篇文章做 4 项 LLM 增强。
 
+    顺序:classify(黑名单 gate) → format → image → commentary
+    - classify 命中 blocklist → 整篇文章 DELETE,后续任务直接 return
+    - 任一任务失败,只标 status 不影响其他任务
+
     返回 {task: status} 字典(用于日志)。
     """
     async with AsyncSessionLocal() as session:
@@ -240,30 +295,54 @@ async def enrich_article(article_id: int) -> dict[str, str]:
         if not art:
             return {}
 
-        # 4 个任务(互不影响);format / classify / commentary 是 chat,image 是 image
+        # === 1) classify(黑名单 gate,优先执行)===
-        # 串行执行(已经过 client 内部 Semaphore),但每个 try/except 独立
+        blocklist = _merge_blocklist(setting, art.source if art.source_id else None)
-        tasks: list[tuple[str, Any]] = [
+        try:
-            ("format", _enrich_format(art, setting, client)),
+            drop, cats = await _enrich_classify(art, setting, client, blocklist)
-            ("classify", _enrich_classify(art, setting, client)),
+            art.classify_status = "ok"
-            ("image", _enrich_image(art, setting, client)),
+            if cats:
-            ("commentary", _enrich_commentary(art, setting, client)),
+                art.category = ",".join(cats)[:64] or None
-        ]
+            if drop:
-        for name, coro in tasks:
+                # 命中 blocklist → 删文章,后续 3 步全跳
-            try:
+                logger.info(
-                await coro
+                    "enrich_article id=%s dropped (blocklist hit, cats=%s, blocklist=%s)",
-                results[name] = "ok"
+                    article_id, cats, blocklist,
-            except Exception as e:
+                )
-                logger.exception("enrich %s failed for article %s: %s", name, article_id, e)
+                await session.delete(art)
-                results[name] = f"failed:{type(e).__name__}"
+                await session.commit()
-                # 标 status
+                return {"classify": "dropped", "format": "skipped", "image": "skipped", "commentary": "skipped"}
-                if name == "format":
+        except Exception as e:
-                    art.format_status = "failed"
+            logger.exception("enrich classify failed for article %s: %s", article_id, e)
-                elif name == "classify":
+            art.classify_status = "failed"
-                    art.classify_status = "failed"
+            results["classify"] = f"failed:{type(e).__name__}"
-                elif name == "image":
+            # classify 失败也继续(format/image/commentary 还能跑)
-                    art.image_ai_status = "failed"
+
-                elif name == "commentary":
+        # === 2) format ===
-                    art.commentary_status = "failed"
+        try:
+            await _enrich_format(art, setting, client)
+            results["format"] = "ok"
+        except Exception as e:
+            logger.exception("enrich format failed for article %s: %s", article_id, e)
+            art.format_status = "failed"
+            results["format"] = f"failed:{type(e).__name__}"
+
+        # === 3) image ===
+        try:
+            await _enrich_image(art, setting, client)
+            results["image"] = "ok"
+        except Exception as e:
+            logger.exception("enrich image failed for article %s: %s", article_id, e)
+            art.image_ai_status = "failed"
+            results["image"] = f"failed:{type(e).__name__}"
+
+        # === 4) commentary ===
+        try:
+            await _enrich_commentary(art, setting, client)
+            results["commentary"] = "ok"
+        except Exception as e:
+            logger.exception("enrich commentary failed for article %s: %s", article_id, e)
+            art.commentary_status = "failed"
+            results["commentary"] = f"failed:{type(e).__name__}"
 
         await session.commit()
     logger.info("enrich_article id=%s: %s", article_id, results)