feat: hashtag 列表显示按共现计算的相关 tag

- 新增 /api/hashtag/related：从作品描述实时解析 hashtag，按「同一作品里一起出现」计算相关 tag - 打分取子话题（候选多与目标同现）与上位话题（目标多与热门候选同现）两者较大值，过滤 cos、二次元等万金油 tag 与平台活动、纯数字 tag - 播放页侧栏的 hashtag 列表顶部显示相关 tag，点击在侧栏内预览（不打断当前播放），在预览里点作品才切换到该 tag 的列表 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-09 01:18:04 +08:00
parent b84fcec77b
commit 2391abdeea
7 changed files with 279 additions and 7 deletions
@@ -0,0 +1,114 @@
+import {
+  HASHTAG_EXTRACT_PG,
+  META_HASHTAG_PG,
+  META_HASHTAGS,
+} from "@/lib/hashtag";
+import { prisma } from "@/lib/prisma";
+
+export type RelatedHashtag = {
+  name: string;
+  /** 和目标 tag 一起出现的作品数 */
+  together: number;
+  /** 这个 tag 自己的作品数 */
+  count: number;
+};
+
+export type RelatedHashtagsResult = {
+  name: string;
+  count: number;
+  related: RelatedHashtag[];
+};
+
+const CACHE_MS = 60_000;
+const cache = new Map<string, { at: number; value: RelatedHashtagsResult }>();
+
+/**
+ * 按「同一个作品描述里一起出现」计算相关 hashtag（共现图），每次实时从描述里解析，不需要额外的表。
+ *
+ * 打分取两个方向里较大的那个：
+ * - 子话题：候选 tag 大多和目标一起出现（together / 候选数），乘 ln(1+together) 防止只出现一次的巧合排到前面；
+ *   例：#明日方舟 → #明日方舟七周年、#特蕾西娅
+ * - 上位话题：目标大多和候选一起出现（together / 目标数），乘 ln(1+候选数)，让热门的上位 tag 排前面；
+ *   例：#圣聆初雪（只有 2 个作品）→ #明日方舟
+ * cos、二次元这类「万金油」tag 与平台活动 tag 不参与推荐。
+ */
+export async function relatedHashtags(
+  rawName: string,
+  limit = 12,
+): Promise<RelatedHashtagsResult> {
+  const name = rawName.toLowerCase();
+  const key = `${name}:${limit}`;
+  const hit = cache.get(key);
+  if (hit && Date.now() - hit.at < CACHE_MS) return hit.value;
+
+  const rows = await prisma.$queryRaw<
+    { name: string; together: number; count: number; target: number }[]
+  >`
+    WITH ht AS (
+      SELECT DISTINCT w.aweme_id, lower(m[1]) AS tag
+      FROM (
+        SELECT aweme_id, "desc" FROM "Video"
+        UNION ALL
+        SELECT aweme_id, "desc" FROM "ImagePost"
+      ) w,
+      regexp_matches(w."desc", ${HASHTAG_EXTRACT_PG}, 'g') m
+    ),
+    target AS (SELECT aweme_id FROM ht WHERE tag = ${name}),
+    target_count AS (SELECT count(*)::int AS c FROM target),
+    co AS (
+      SELECT ht.tag, count(*)::int AS together
+      FROM ht JOIN target USING (aweme_id)
+      WHERE ht.tag <> ${name}
+      GROUP BY ht.tag
+    ),
+    counts AS (
+      SELECT tag, count(*)::int AS c FROM ht
+      WHERE tag IN (SELECT tag FROM co)
+      GROUP BY tag
+    )
+    SELECT
+      co.tag AS name,
+      co.together,
+      counts.c AS count,
+      tc.c AS target
+    FROM co
+    JOIN counts USING (tag)
+    CROSS JOIN target_count tc
+    WHERE co.together >= CASE WHEN tc.c >= 4 THEN 2 ELSE 1 END
+      AND co.tag <> ALL (${META_HASHTAGS}::text[])
+      AND co.tag !~ ${META_HASHTAG_PG}
+    ORDER BY GREATEST(
+      co.together::float8 / counts.c * ln(1 + co.together),
+      co.together::float8 / tc.c * ln(1 + counts.c)
+    ) DESC, co.together DESC, co.tag
+    LIMIT ${limit}
+  `;
+
+  let count = rows[0]?.target;
+  if (count === undefined) {
+    // 没有相关 tag 时单独数一下目标 tag 的作品数
+    const [{ c }] = await prisma.$queryRaw<{ c: number }[]>`
+      SELECT count(DISTINCT w.aweme_id)::int AS c
+      FROM (
+        SELECT aweme_id, "desc" FROM "Video"
+        UNION ALL
+        SELECT aweme_id, "desc" FROM "ImagePost"
+      ) w,
+      regexp_matches(w."desc", ${HASHTAG_EXTRACT_PG}, 'g') m
+      WHERE lower(m[1]) = ${name}
+    `;
+    count = c;
+  }
+
+  const value: RelatedHashtagsResult = {
+    name,
+    count,
+    related: rows.map(({ name, together, count }) => ({
+      name,
+      together,
+      count,
+    })),
+  };
+  cache.set(key, { at: Date.now(), value });
+  return value;
+}