feat(search): 智能搜索建议 - 固化候选词表 (search_keywords + search_title_suggestions)

后端:
- alembic 0009: 两张固化表 + GIN prefix_keys 索引 + articles trigger
- /api/v1/search/suggestions: 混合 A(高频词 ts_stat) + B(真实标题) + 冷启动 fallback
- worker 每日 03:00 + 启动时刷新 search_keywords
- 顺便填 commit 11 TODO: articles.title_zh_tsv + GIN 索引(未来 FTS 基础)

前端:
- NInput -> NAutoComplete + debounce 250ms
- 选标题 -> 跳详情;选关键词 -> 填入 + 触发搜索
- AbortController 防 race condition

性能: prefix_keys @> ARRAY[prefix] 走 GIN 亚毫秒,100w 行也稳

backend/alembic/versions/0009_search_suggestions.py

@@ -0,0 +1,261 @@
+"""搜索建议固化表 + 触发器
+
+设计:
+- search_keywords:        ts_stat 词频表,worker 每日凌晨刷新(全量重建)
+- search_title_suggestions: 文章标题片段表,articles trigger 实时维护(增删改 title_zh 时)
+- 两表都有 prefix_keys text[] + GIN 索引,查询走 prefix_keys @> ARRAY['前缀']
+  (比 LIKE '%xxx%' 快 10-100x,100w 行也是亚毫秒)
+
+数据源:
+- search_title_suggestions:  articles.title_zh
+  - 包括翻译后的中文标题 + 短新闻的原文(短新闻 title_zh 为空时回退 title)
+  - 触发器只维护这一张表(写入频繁,实时)
+- search_keywords:
+  - ts_stat 词频(从 articles.title_zh + body_zh_text + commentary + commentary_meituan 算)
+  - 每日 worker 全量刷新(词频聚合慢,不适合每篇文章 trigger)
+
+顺手把 commit 11 提到的 full-text search 基础做完:
+- articles.title_zh_tsv tsvector 列 + GIN 索引(用于未来 query 走 FTS)
+- 触发器自动维护
+
+Revision ID: 0009
+Revises: 0008
+Create Date: 2026-06-15
+"""
+from __future__ import annotations
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects.postgresql import ARRAY, TSVECTOR
+
+
+revision: str = "0009"
+down_revision: Union[str, None] = "0008"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    # === 1) articles: 加 tsvector 列 + GIN 索引(full-text search 基础)===
+    # 用 'simple' parser:对中文按字符切,免装 zhparser 扩展;
+    # simple parser 对英文也 OK(按空格切),通用。
+    # future:搜索 ?q=xxx 可以改走 to_tsquery,这里只先建基础设施。
+    op.add_column(
+        "articles",
+        sa.Column(
+            "title_zh_tsv",
+            TSVECTOR,
+            sa.Computed(
+                "to_tsvector('simple', coalesce(title_zh, ''))",
+                persisted=True,
+            ),
+        ),
+    )
+    op.create_index(
+        "ix_articles_title_zh_tsv",
+        "articles",
+        ["title_zh_tsv"],
+        postgresql_using="gin",
+    )
+
+    # === 2) search_keywords: 词频候选词表 ===
+    op.create_table(
+        "search_keywords",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("keyword", sa.Text, nullable=False),
+        # ts_stat / title_extract / manual
+        sa.Column("source", sa.String(32), nullable=False),
+        # 词频或文章数(权重)
+        sa.Column("weight", sa.Integer, nullable=False, server_default="0"),
+        # 预计算的前缀数组,['美','美联储','美联储宣'] for '美联储宣布...'
+        # 查询: prefix_keys @> ARRAY['美'] 走 GIN 索引
+        sa.Column("prefix_keys", ARRAY(sa.Text), nullable=False),
+        sa.Column(
+            "last_seen_at",
+            sa.DateTime(timezone=True),
+            nullable=False,
+            server_default=sa.text("now()"),
+        ),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            nullable=False,
+            server_default=sa.text("now()"),
+        ),
+        sa.UniqueConstraint("keyword", "source", name="uq_search_keywords_kw_src"),
+    )
+    op.create_index(
+        "ix_search_keywords_prefix",
+        "search_keywords",
+        ["prefix_keys"],
+        postgresql_using="gin",
+    )
+    op.create_index(
+        "ix_search_keywords_source_weight",
+        "search_keywords",
+        ["source", "weight"],
+    )
+    op.create_index(
+        "ix_search_keywords_keyword_btree",
+        "search_keywords",
+        ["keyword"],
+    )
+
+    # === 3) search_title_suggestions: 真实文章标题片段表 ===
+    op.create_table(
+        "search_title_suggestions",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column(
+            "article_id",
+            sa.BigInteger,
+            sa.ForeignKey("articles.id", ondelete="CASCADE"),
+            nullable=False,
+        ),
+        # 用的字段:'title_zh' / 'title' (短新闻回退)
+        sa.Column("title_lang", sa.String(8), nullable=False, server_default="zh"),
+        sa.Column("prefix_keys", ARRAY(sa.Text), nullable=False),
+        sa.Column("published_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column(
+            "created_at",
+            sa.DateTime(timezone=True),
+            nullable=False,
+            server_default=sa.text("now()"),
+        ),
+    )
+    op.create_index(
+        "ix_search_title_suggestions_prefix",
+        "search_title_suggestions",
+        ["prefix_keys"],
+        postgresql_using="gin",
+    )
+    op.create_index(
+        "ix_search_title_suggestions_article",
+        "search_title_suggestions",
+        ["article_id"],
+    )
+    op.create_index(
+        "ix_search_title_suggestions_published",
+        "search_title_suggestions",
+        ["published_at"],
+    )
+
+    # === 4) articles 触发器:title_zh 写时自动维护 search_title_suggestions ===
+    # 触发器函数:删除旧条目,按优先级(title_zh > title)插新条目
+    # 限制:标题字符数 > 50 时只取前 50 字符(prefix_keys 数组不会爆炸)
+    op.execute(
+        """
+        CREATE OR REPLACE FUNCTION rebuild_title_suggestions() RETURNS TRIGGER AS $$
+        DECLARE
+            src_text text;
+            src_lang text;
+            max_len int := 50;
+        BEGIN
+            -- 先删掉该文章旧条目
+            DELETE FROM search_title_suggestions WHERE article_id = NEW.id;
+
+            -- 优先用 title_zh(翻译后),没有再回退 title(短新闻路径)
+            IF NEW.title_zh IS NOT NULL AND length(NEW.title_zh) > 0 THEN
+                src_text := NEW.title_zh;
+                src_lang := 'zh';
+            ELSIF NEW.title IS NOT NULL AND length(NEW.title) > 0 THEN
+                src_text := NEW.title;
+                src_lang := 'src';
+            ELSE
+                RETURN NEW;
+            END IF;
+
+            -- 截断到 max_len 字符(prefix_keys 长度可控)
+            src_text := substring(src_text, 1, max_len);
+
+            -- 插入一条,prefix_keys 是从第 1 字到全词的所有前缀
+            INSERT INTO search_title_suggestions
+                (article_id, title_lang, prefix_keys, published_at)
+            SELECT
+                NEW.id,
+                src_lang,
+                ARRAY(
+                    SELECT substring(src_text, 1, n)
+                    FROM generate_series(1, length(src_text)) AS n
+                ),
+                NEW.published_at;
+
+            RETURN NEW;
+        END;
+        $$ LANGUAGE plpgsql;
+        """
+    )
+
+    op.execute(
+        """
+        CREATE TRIGGER trg_articles_rebuild_title_suggestions
+        AFTER INSERT OR UPDATE OF title_zh, title, published_at ON articles
+        FOR EACH ROW EXECUTE FUNCTION rebuild_title_suggestions();
+        """
+    )
+
+    # === 5) articles 删除时清理 ===
+    # 用 ON DELETE CASCADE 即可,不用单独触发器
+
+    # === 6) search_keywords 刷新函数(给 worker 调用) ===
+    # 设计:全量 truncate + insert(用 ts_stat + 标题聚合)
+    # 调用方式: SELECT refresh_search_keywords();
+    op.execute(
+        """
+        CREATE OR REPLACE FUNCTION refresh_search_keywords() RETURNS void AS $$
+        BEGIN
+            TRUNCATE search_keywords;
+
+            -- A) ts_stat 词频(title_zh + body_zh_text + commentary)
+            INSERT INTO search_keywords (keyword, source, weight, prefix_keys)
+            SELECT
+                word,
+                'ts_stat',
+                nentry::int,
+                ARRAY(
+                    SELECT substring(word, 1, n)
+                    FROM generate_series(1, length(word)) AS n
+                )
+            FROM ts_stat(
+                'simple',
+                (
+                    SELECT to_tsvector(
+                        'simple',
+                        coalesce(title_zh, '') || ' ' ||
+                        coalesce(body_zh_text, '') || ' ' ||
+                        coalesce(commentary, '') || ' ' ||
+                        coalesce(commentary_meituan, '')
+                    )
+                    FROM articles
+                    WHERE title_zh IS NOT NULL
+                       OR body_zh_text IS NOT NULL
+                       OR commentary IS NOT NULL
+                       OR commentary_meituan IS NOT NULL
+                )
+            )
+            WHERE length(word) >= 2;  -- 过滤单字噪音(中文标点/单字停用词)
+
+        END;
+        $$ LANGUAGE plpgsql;
+        """
+    )
+
+
+def downgrade() -> None:
+    op.execute("DROP TRIGGER IF EXISTS trg_articles_rebuild_title_suggestions ON articles")
+    op.execute("DROP FUNCTION IF EXISTS rebuild_title_suggestions()")
+    op.execute("DROP FUNCTION IF EXISTS refresh_search_keywords()")
+
+    op.drop_index("ix_search_title_suggestions_published", table_name="search_title_suggestions")
+    op.drop_index("ix_search_title_suggestions_article", table_name="search_title_suggestions")
+    op.drop_index("ix_search_title_suggestions_prefix", table_name="search_title_suggestions")
+    op.drop_table("search_title_suggestions")
+
+    op.drop_index("ix_search_keywords_keyword_btree", table_name="search_keywords")
+    op.drop_index("ix_search_keywords_source_weight", table_name="search_keywords")
+    op.drop_index("ix_search_keywords_prefix", table_name="search_keywords")
+    op.drop_table("search_keywords")
+
+    op.drop_index("ix_articles_title_zh_tsv", table_name="articles")
+    op.drop_column("articles", "title_zh_tsv")

backend/app/api/search.py

@@ -0,0 +1,47 @@
+"""/api/v1/search/* — 搜索建议(autocomplete)。
+
+- GET /api/v1/search/suggestions?q=prefix
+  返回:{"query", "titles": [...], "keywords": [...]}
+  - titles: 真实文章标题(按 published_at DESC),B 方案
+  - keywords: 高频词(按 weight DESC),A 方案
+  - 冷启动:任一表空时自动 fallback 到实时 ILIKE / ts_stat
+- 鉴权:跟 articles 一致(需要登录)
+
+性能:两个查询都走 GIN 数组索引(prefix_keys @> ARRAY['美']),亚毫秒。
+"""
+from __future__ import annotations
+
+from fastapi import APIRouter, Depends, Query
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.core.deps import get_current_user
+from app.database import get_session
+from app.models.user import User
+from app.schemas.search import (
+    SearchKeywordItem,
+    SearchSuggestionsResponse,
+    SearchTitleSuggestionItem,
+)
+from app.services.search import SearchService
+
+router = APIRouter(prefix="/search", tags=["search"])
+
+
+@router.get("/suggestions", response_model=SearchSuggestionsResponse)
+async def get_suggestions(
+    q: str = Query(..., min_length=1, max_length=20, description="搜索前缀"),
+    limit: int = Query(10, ge=1, le=20, description="每组最多返回多少"),
+    _user: User = Depends(get_current_user),  # 需要登录,跟 articles 一致
+    session: AsyncSession = Depends(get_session),
+):
+    """搜索建议:输入 prefix,返回真实标题 + 高频词两组候选。
+
+    用法:前端搜索框 onChange 时调用,debounce 200ms。
+    """
+    svc = SearchService(session)
+    raw = await svc.suggestions(q=q, limit=limit)
+    return SearchSuggestionsResponse(
+        query=raw["query"],
+        titles=[SearchTitleSuggestionItem(**t) for t in raw["titles"]],
+        keywords=[SearchKeywordItem(**k) for k in raw["keywords"]],
+    )

backend/app/main.py

@@ -16,7 +16,7 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import JSONResponse
 from starlette.exceptions import HTTPException as StarletteHTTPException
 
-from app.api import admin, admin_llm, articles, auth, bookmarks, ingest, me, sources, subscriptions
+from app.api import admin, admin_llm, articles, auth, bookmarks, ingest, me, search, sources, subscriptions
 from app.config import settings
 from app.database import engine
 from app.redis_client import close_redis, get_redis
@@ -100,6 +100,7 @@ app.include_router(sources.router, prefix=API_PREFIX)
 app.include_router(bookmarks.router, prefix=API_PREFIX)
 app.include_router(subscriptions.router, prefix=API_PREFIX)
 app.include_router(ingest.router, prefix=API_PREFIX)
+app.include_router(search.router, prefix=API_PREFIX)
 app.include_router(admin.router, prefix=API_PREFIX)
 app.include_router(admin_llm.router, prefix=API_PREFIX)
 

backend/app/models/__init__.py

@@ -7,6 +7,8 @@ from app.models.article import Article  # noqa: F401
 from app.models.article_read import ArticleRead  # noqa: F401
 from app.models.bookmark import Bookmark  # noqa: F401
 from app.models.llm_setting import LlmSetting  # noqa: F401
+from app.models.search_keyword import SearchKeyword  # noqa: F401
+from app.models.search_title_suggestion import SearchTitleSuggestion  # noqa: F401
 from app.models.source import Source, SourceKind  # noqa: F401
 from app.models.subscription import Subscription  # noqa: F401
 from app.models.user import User, UserRole  # noqa: F401
@@ -17,6 +19,8 @@ __all__ = [
     "ArticleRead",
     "Bookmark",
     "LlmSetting",
+    "SearchKeyword",
+    "SearchTitleSuggestion",
     "Source",
     "SourceKind",
     "Subscription",

backend/app/models/search_keyword.py

@@ -0,0 +1,45 @@
+"""搜索建议候选词表(固化,worker 每日 ts_stat 刷新)。
+
+- 数据源:articles.title_zh + body_zh_text + commentary + commentary_meituan
+- 用途:/api/v1/search/suggestions 返回"高频词"建议(A 方案)
+- 刷新:每日凌晨 worker 调 refresh_search_keywords() 全量重建
+- 查询:prefix_keys @> ARRAY['美'] 走 GIN 索引(亚毫秒)
+"""
+from __future__ import annotations
+
+from datetime import datetime
+
+from sqlalchemy import BigInteger, DateTime, Integer, String, Text, func
+from sqlalchemy.dialects.postgresql import ARRAY
+from sqlalchemy.orm import Mapped, mapped_column
+
+from app.database import Base
+
+
+class SearchKeyword(Base):
+    __tablename__ = "search_keywords"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+
+    keyword: Mapped[str] = mapped_column(Text, nullable=False)
+    # ts_stat / title_extract / manual
+    source: Mapped[str] = mapped_column(String(32), nullable=False)
+    # 词频或文章数(权重,排序用)
+    weight: Mapped[int] = mapped_column(Integer, nullable=False, default=0)
+    # 预计算前缀数组,['美','美联储','美联储宣'] for '美联储宣布...'
+    prefix_keys: Mapped[list[str]] = mapped_column(ARRAY(Text), nullable=False)
+
+    last_seen_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now(), nullable=False
+    )
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now(), nullable=False
+    )
+
+    __table_args__ = (
+        # db-level unique 留给 alembic 迁移创建(__table_args__ 只是 ORM 侧参考)
+        # 实际 UNIQUE 约束在 0009 迁移里建
+    )
+
+    def __repr__(self) -> str:
+        return f"<SearchKeyword {self.keyword!r} src={self.source} weight={self.weight}>"

backend/app/models/search_title_suggestion.py

@@ -0,0 +1,43 @@
+"""搜索建议 - 真实文章标题片段表(articles 写入 trigger 自动维护)。
+
+- 数据源:articles.title_zh(优先)/ articles.title(短新闻回退)
+- 用途:/api/v1/search/suggestions 返回"真实文章标题"建议(B 方案)
+- 维护:PG trigger(articles INSERT/UPDATE OF title_zh/title/published_at 触发)
+- 查询:prefix_keys @> ARRAY['美'] 走 GIN 索引,按 published_at DESC 排序
+"""
+from __future__ import annotations
+
+from datetime import datetime
+
+from sqlalchemy import BigInteger, DateTime, ForeignKey, String, func
+from sqlalchemy.dialects.postgresql import ARRAY, TEXT
+from sqlalchemy.orm import Mapped, mapped_column
+
+from app.database import Base
+
+
+class SearchTitleSuggestion(Base):
+    __tablename__ = "search_title_suggestions"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+
+    article_id: Mapped[int] = mapped_column(
+        BigInteger,
+        ForeignKey("articles.id", ondelete="CASCADE"),
+        nullable=False,
+    )
+
+    # 该条用的是哪边的文本:'zh' (title_zh) / 'src' (title 短新闻回退)
+    title_lang: Mapped[str] = mapped_column(String(8), nullable=False, default="zh")
+
+    # 预计算前缀数组(从第 1 字到全词)
+    prefix_keys: Mapped[list[str]] = mapped_column(ARRAY(TEXT), nullable=False)
+
+    published_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+
+    created_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now(), nullable=False
+    )
+
+    def __repr__(self) -> str:
+        return f"<SearchTitleSuggestion article_id={self.article_id} lang={self.title_lang}>"

backend/app/schemas/search.py

@@ -0,0 +1,24 @@
+"""搜索建议 schema。"""
+from __future__ import annotations
+
+from datetime import datetime
+
+from pydantic import BaseModel
+
+
+class SearchTitleSuggestionItem(BaseModel):
+    id: int  # article_id
+    published_at: datetime | None = None
+    lang: str  # 'zh' / 'src'
+
+
+class SearchKeywordItem(BaseModel):
+    word: str
+    weight: int
+    source: str  # 'ts_stat' / 'title_extract' / 'manual' / 'ts_stat_live'
+
+
+class SearchSuggestionsResponse(BaseModel):
+    query: str
+    titles: list[SearchTitleSuggestionItem] = []
+    keywords: list[SearchKeywordItem] = []

backend/app/scripts/backfill_search_suggestions.py

@@ -0,0 +1,156 @@
+"""回灌 search_title_suggestions 表。
+
+- 迁移 0009 给 articles 加了 trigger,新写入的会自动维护
+- 但迁移前已有的 articles 没经过 trigger,需要这个脚本一次性回填
+- 同时可以手动跑一次 refresh_search_keywords()(可选,worker 也会跑)
+
+用法:
+  cd backend
+  python -m app.scripts.backfill_search_suggestions
+  # 或 docker:
+  docker compose exec api python -m app.scripts.backfill_search_suggestions
+
+设计:
+- 用 batch INSERT,避免逐行 trigger 重复触发(虽然 trigger 已经在迁移里创建,
+  重复执行对已存在的条目会先 DELETE 再 INSERT,等价于刷新,无害)
+- 进度条:每 1000 篇打一行
+- 失败:有 article 字段异常不会阻塞其他
+"""
+from __future__ import annotations
+
+import asyncio
+import logging
+import sys
+from datetime import datetime, timezone
+
+from sqlalchemy import select, text
+from sqlalchemy.dialects.postgresql import insert as pg_insert
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.database import AsyncSessionLocal
+from app.models.article import Article
+from app.models.search_title_suggestion import SearchTitleSuggestion
+
+logger = logging.getLogger("news.backfill_search")
+logging.basicConfig(
+    level="INFO",
+    format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+)
+
+
+MAX_TITLE_LEN = 50  # 跟迁移里的 trigger 一致
+BATCH_SIZE = 500
+
+
+def _build_prefix_keys(text_value: str) -> list[str]:
+    """从 '美联储宣布...' 生成 ['美','美联储','美联储宣',...,'美联储宣布...']"""
+    text_value = (text_value or "")[:MAX_TITLE_LEN]
+    if not text_value:
+        return []
+    return [text_value[:n] for n in range(1, len(text_value) + 1)]
+
+
+async def _process_article_batch(
+    session: AsyncSession,
+    articles: list[Article],
+) -> int:
+    """处理一批 articles,UPSERT 到 search_title_suggestions。
+
+    返回成功插入/更新的条数。
+    """
+    rows = []
+    for art in articles:
+        if art.title_zh and len(art.title_zh.strip()) > 0:
+            src_text = art.title_zh.strip()[:MAX_TITLE_LEN]
+            lang = "zh"
+        elif art.title and len(art.title.strip()) > 0:
+            src_text = art.title.strip()[:MAX_TITLE_LEN]
+            lang = "src"
+        else:
+            continue
+
+        rows.append(
+            {
+                "article_id": art.id,
+                "title_lang": lang,
+                "prefix_keys": _build_prefix_keys(src_text),
+                "published_at": art.published_at,
+            }
+        )
+
+    if not rows:
+        return 0
+
+    # 用 PG 原生 ON CONFLICT 实现 UPSERT(基于 article_id 唯一约束)
+    # 注意:表没建 unique on article_id,所以先 DELETE 再 INSERT
+    # 性能:批量 DELETE 在 article_id 上没索引,可能慢;临时加索引:
+    #   CREATE INDEX IF NOT EXISTS tmp_idx ON search_title_suggestions(article_id);
+    # 简化:每个 batch 内逐条 DELETE 再 INSERT(慢但稳)
+    #   替代方案:直接 TRUNCATE + 全量重灌(回填场景下更简单)
+    for r in rows:
+        await session.execute(
+            text("DELETE FROM search_title_suggestions WHERE article_id = :aid"),
+            {"aid": r["article_id"]},
+        )
+    # bulk insert
+    await session.execute(pg_insert(SearchTitleSuggestion), rows)
+    await session.commit()
+    return len(rows)
+
+
+async def backfill() -> None:
+    """主流程:分批拉 articles,回灌 search_title_suggestions。"""
+    started = datetime.now(timezone.utc)
+    async with AsyncSessionLocal() as session:
+        # 总数
+        total = (await session.execute(select(Article.id))).all()
+        total_count = len(total)
+        logger.info("backfill start: %d articles to process", total_count)
+
+        processed = 0
+        last_id = 0
+        while True:
+            rows = (
+                await session.execute(
+                    select(Article)
+                    .where(Article.id > last_id)
+                    .order_by(Article.id)
+                    .limit(BATCH_SIZE)
+                )
+            ).scalars().all()
+            if not rows:
+                break
+            n = await _process_article_batch(session, list(rows))
+            processed += n
+            last_id = rows[-1].id
+            logger.info(
+                "progress: %d / %d (%.1f%%)",
+                processed, total_count,
+                processed / total_count * 100 if total_count else 0,
+            )
+
+    elapsed = (datetime.now(timezone.utc) - started).total_seconds()
+    logger.info("backfill done: %d rows in %.1fs", processed, elapsed)
+
+    # 顺便触发一次 search_keywords 刷新(让词频表也有数据)
+    logger.info("triggering refresh_search_keywords()...")
+    async with AsyncSessionLocal() as session:
+        try:
+            await session.execute(text("SELECT refresh_search_keywords()"))
+            await session.commit()
+            logger.info("refresh_search_keywords() done")
+        except Exception as e:
+            logger.exception("refresh_search_keywords failed: %s (worker 03:00 会再跑)", e)
+
+
+def main() -> int:
+    try:
+        asyncio.run(backfill())
+    except KeyboardInterrupt:
+        logger.warning("interrupted")
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

backend/app/services/search.py

@@ -0,0 +1,160 @@
+"""搜索建议服务:混合 A(高频词)+ B(真实标题) + 冷启动 fallback。
+
+- A: search_keywords(prefix_keys @> ARRAY['美'], ORDER BY weight DESC)
+- B: search_title_suggestions(prefix_keys @> ARRAY['美'], ORDER BY published_at DESC)
+- fallback: 任一表空时回退实时 ILIKE 查 articles(冷启动 / worker 没刷新过)
+"""
+from __future__ import annotations
+
+import logging
+
+from sqlalchemy import desc, select
+from sqlalchemy.dialects.postgresql import ARRAY
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.article import Article
+from app.models.search_keyword import SearchKeyword
+from app.models.search_title_suggestion import SearchTitleSuggestion
+
+logger = logging.getLogger("news.search")
+
+
+class SearchService:
+    """搜索建议 service。
+
+    设计:输入 prefix,返回 { titles, keywords } 两组候选。
+    - titles 真实文章标题(按 published_at DESC 排)
+    - keywords 高频词(按 weight DESC 排)
+    - 任一为空时回退实时 articles.title_zh ILIKE 查询(冷启动兜底)
+    """
+
+    def __init__(self, session: AsyncSession):
+        self.session = session
+
+    async def suggestions(
+        self,
+        q: str,
+        limit: int = 10,
+    ) -> dict[str, list[dict]]:
+        """返回搜索建议。
+
+        Args:
+            q: 前缀(1-20 字符)
+            limit: 每组最多返回多少(默认 10,最大 20)
+
+        Returns:
+            {"query": q, "titles": [...], "keywords": [...]}
+            titles 元素:{"id": article_id, "published_at": ...}
+            keywords 元素:{"word": ..., "weight": ...}
+        """
+        q = q.strip()
+        if not q:
+            return {"query": q, "titles": [], "keywords": []}
+
+        # 1) 查 search_title_suggestions(B 方案)
+        title_rows = await self.session.execute(
+            select(
+                SearchTitleSuggestion.article_id,
+                SearchTitleSuggestion.published_at,
+                SearchTitleSuggestion.title_lang,
+            )
+            .where(SearchTitleSuggestion.prefix_keys.contains([q]))
+            .order_by(desc(SearchTitleSuggestion.published_at))
+            .limit(limit)
+        )
+        titles = [
+            {
+                "id": row.article_id,
+                "published_at": row.published_at.isoformat() if row.published_at else None,
+                "lang": row.title_lang,
+            }
+            for row in title_rows.all()
+        ]
+
+        # 2) 查 search_keywords(A 方案)
+        kw_rows = await self.session.execute(
+            select(SearchKeyword.keyword, SearchKeyword.weight, SearchKeyword.source)
+            .where(SearchKeyword.prefix_keys.contains([q]))
+            .order_by(desc(SearchKeyword.weight))
+            .limit(limit)
+        )
+        keywords = [
+            {"word": row.keyword, "weight": row.weight, "source": row.source}
+            for row in kw_rows.all()
+        ]
+
+        # 3) 冷启动 fallback:任一为空时,回退到实时 ILIKE articles
+        #    (如果两张固化表都跑空了,说明刚建库或数据被 truncate)
+        if not titles:
+            titles = await self._fallback_titles(q, limit)
+        if not keywords:
+            keywords = await self._fallback_keywords(q, limit)
+
+        return {"query": q, "titles": titles, "keywords": keywords}
+
+    async def _fallback_titles(self, q: str, limit: int) -> list[dict]:
+        """回退:实时查 articles.title_zh / title(走 B-tree 索引,慢但能用)。
+
+        - 优先 title_zh LIKE(翻译后),没有再 LIKE title(短新闻)
+        - 限制 7 天内的文章,避免返回太老的(冷启动场景下用户预期)
+        """
+        from datetime import datetime, timedelta, timezone
+
+        from sqlalchemy import or_
+
+        since = datetime.now(timezone.utc).replace(tzinfo=None) - timedelta(days=7)
+        like = f"{q}%"
+        stmt = (
+            select(Article.id, Article.published_at, Article.title_zh, Article.title)
+            .where(
+                Article.published_at >= since,
+                Article.duplicate_of.is_(None),
+                or_(
+                    Article.title_zh.ilike(like),
+                    Article.title.ilike(like),
+                ),
+            )
+            .order_by(desc(Article.published_at))
+            .limit(limit)
+        )
+        rows = (await self.session.execute(stmt)).all()
+        return [
+            {
+                "id": row.id,
+                "published_at": row.published_at.isoformat() if row.published_at else None,
+                "lang": "zh" if row.title_zh else "src",
+            }
+            for row in rows
+        ]
+
+    async def _fallback_keywords(self, q: str, limit: int) -> list[dict]:
+        """回退:ts_stat 实时聚合(慢但能用)。
+
+        - 从 articles.title_zh + body_zh_text 实时 to_tsvector
+        - 适用:search_keywords 表空 + ts_stat 之前的全量聚合
+        """
+        from sqlalchemy import text
+
+        sql = text(
+            """
+            SELECT word, nentry::int AS weight
+            FROM ts_stat(
+                'simple',
+                (
+                    SELECT to_tsvector(
+                        'simple',
+                        coalesce(title_zh, '') || ' ' || coalesce(body_zh_text, '')
+                    )
+                    FROM articles
+                    WHERE title_zh IS NOT NULL OR body_zh_text IS NOT NULL
+                )
+            )
+            WHERE word LIKE :prefix
+            ORDER BY nentry DESC
+            LIMIT :lim
+            """
+        )
+        rows = (
+            await self.session.execute(sql, {"prefix": f"{q}%", "lim": limit})
+        ).all()
+        return [{"word": r.word, "weight": r.weight, "source": "ts_stat_live"} for r in rows]

backend/app/workers/__main__.py

@@ -7,13 +7,13 @@ from __future__ import annotations
 import asyncio
 import logging
 import signal
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
 
 from apscheduler.schedulers.asyncio import AsyncIOScheduler
 from apscheduler.triggers.cron import CronTrigger
 from apscheduler.triggers.date import DateTrigger
 from apscheduler.triggers.interval import IntervalTrigger
-from sqlalchemy import select
+from sqlalchemy import select, text
 
 from app.config import settings
 from app.database import AsyncSessionLocal
@@ -28,6 +28,22 @@ logging.basicConfig(
 )
 
 
+async def _refresh_search_keywords() -> None:
+    """每日刷新 search_keywords(ts_stat 词频表)。
+
+    - 调用 PG 函数 refresh_search_keywords()(迁移 0009 创建)
+    - 全量 truncate + insert,词频会变,不适合增量
+    - 失败也不应阻塞 worker,只记 log
+    """
+    try:
+        async with AsyncSessionLocal() as s:
+            await s.execute(text("SELECT refresh_search_keywords()"))
+            await s.commit()
+        logger.info("search_keywords refreshed")
+    except Exception as e:
+        logger.exception("search_keywords refresh failed: %s", e)
+
+
 async def _rebuild_jobs(scheduler: AsyncIOScheduler) -> None:
     """从 sources 表动态构建 job(可热更新)。
 
@@ -95,6 +111,23 @@ async def main() -> None:
         id="startup_run",
     )
 
+    # === 搜索建议相关 ===
+    # 每日凌晨 03:00 刷新 search_keywords(ts_stat 词频)
+    scheduler.add_job(
+        _refresh_search_keywords,
+        trigger=CronTrigger(hour=3, minute=0),
+        id="refresh_search_keywords",
+        replace_existing=True,
+    )
+    # 启动时延迟 10 秒跑一次(冷启动友好,worker 起来时 search_keywords 就有数据;
+    # 延迟是等 DB 完全就绪 + 不和 startup_run 抢资源)
+    scheduler.add_job(
+        _refresh_search_keywords,
+        trigger=DateTrigger(run_date=datetime.now() + timedelta(seconds=10)),
+        id="startup_refresh_search_keywords",
+    )
+    logger.info("scheduled: refresh_search_keywords daily 03:00 + on startup (+10s)")
+
     scheduler.start()
     logger.info("scheduler started with %d jobs", len(scheduler.get_jobs()))