fix(fam-edge): 人物识别结果没同步进事件描述文字

用户反馈：事件卡片上方的人物徽章正确显示"媳妇/爸爸"，但描述文字里还是 "人物B双手叉腰站在客厅中央；人物A在远处厨房"这种原始 uid——两处对不上。根因：rewrite_event_person_names() 只改了 person_list_json / person_appearances_json[].uid 这些结构化字段，从没碰过 description（事件描述）和 videos.summary_json（会话摘要）——这两个是大模型写的自然语言， "人物A"这类 uid 是直接以文字形式嵌在句子里的，结构化字段改名不会带着改。现在 rewrite_event_person_names() 额外对 description/summary_json 做文本替换。替换顺序按 rename_map 的 key 长度降序——uid 可能带 "#2"/"#3" 这类同名冲突后缀，"人物A" 是 "人物A#2" 的前缀，短的先替换会把长的错误地部分命中，长的先替换才对。同时跑了两次一次性回填修历史数据： 1. 对所有已有 person_identity_map 映射的 1265 个视频重新跑一遍 rewrite_event_person_names，把之前只改了徽章、没改描述文字的历史数据补上（events 里"纯人物X不带真名"的还剩 258 条，涉及 94 个视频——都是性别/年龄特征本来就是 unknown 的画面，闭集识别本来就该留空不瞎猜，不是回填没覆盖到，是真的没法从图像信息判断是谁） 2. 找到 9 个从未跑过闭集识别的漏网视频（应该是 person_identifier 功能上线前处理的），重新识别后 1 个（video 2194）成功解析出汤圆+爸爸，其余 8 个同样是性别未知，正确保持未解析新增 2 个测试覆盖 description/summary_json 文本替换，以及带后缀 label 的替换顺序正确性。
2026-08-23 10:50:52 +08:00
parent 63b28663f1
commit 5f7b27bf04
2 changed files with 100 additions and 14 deletions
--- a/fam-edge/src/fam_edge/oracle_db.py
+++ b/fam-edge/src/fam_edge/oracle_db.py
@@ -657,17 +657,36 @@ class OracleDB:

    def rewrite_event_person_names(self, video_id: int, rename_map: Dict[str, str]):
        """按 {当前展示名: 新名} 把该视频全部 events 的 person_list_json /
-        person_appearances_json[].uid，以及 videos.people_json 里的名字替换掉。
-        rename_map 的 key 是"事件数据里当前显示的名字"（可能是原始 uid，也可能是
-        上一轮已经替换过的规范名——纠错场景下就是这种情况）。
+        person_appearances_json[].uid / description 文本，以及 videos.people_json /
+        summary_json 文本里的名字替换掉。rename_map 的 key 是"事件数据里当前显示
+        的名字"（可能是原始 uid，也可能是上一轮已经替换过的规范名——纠错场景就是
+        这种情况）。
+
+        description/summary_json 是大模型写的自然语言描述，"人物A"这类 uid 会
+        直接以文字形式出现在句子里（不只是 person_list_json 这种结构化字段）——
+        只替换结构化字段的话，事件卡片上方的人物徽章会显示正确的规范名，但描述
+        文字里还是"人物A/人物B"，看着两处对不上。这里做纯文本替换来解决。
        """
        if not rename_map:
            return
        now = _now_iso()
+        # 按 key 长度降序替换：uid 可能带 "#2"/"#3" 这类后缀（同名冲突时的派生
+        # label），"人物A" 是 "人物A#2" 的前缀，先替换短的会把长的也错误命中一部分，
+        # 长的先替换就不会被短的抢先破坏。
+        ordered_keys = sorted(rename_map.keys(), key=len, reverse=True)
+
+        def _rewrite_text(text: str) -> str:
+            if not text:
+                return text
+            for old in ordered_keys:
+                if old in text:
+                    text = text.replace(old, rename_map[old])
+            return text
+
        with self._write_lock:
            rows = self._conn.execute(
-                "SELECT id, person_list_json, person_appearances_json FROM events "
-                "WHERE video_id=?", (video_id,)).fetchall()
+                "SELECT id, person_list_json, person_appearances_json, description "
+                "FROM events WHERE video_id=?", (video_id,)).fetchall()
            for r in rows:
                changed = False
                plist = json.loads(r['person_list_json'] or '[]')
@@ -680,23 +699,38 @@ class OracleDB:
                        if isinstance(p, dict) and p.get('uid') in rename_map:
                            p['uid'] = rename_map[p['uid']]
                            changed = True
+                new_desc = _rewrite_text(r['description'])
+                if new_desc != r['description']:
+                    changed = True
                if changed:
                    self._conn.execute(
-                        "UPDATE events SET person_list_json=?, person_appearances_json=? "
-                        "WHERE id=?",
+                        "UPDATE events SET person_list_json=?, person_appearances_json=?, "
+                        "description=? WHERE id=?",
                        (json.dumps(new_plist, ensure_ascii=False),
                         json.dumps(pa, ensure_ascii=False) if pa is not None
                         else r['person_appearances_json'],
+                         new_desc,
                         r['id']))
            vrow = self._conn.execute(
-                "SELECT people_json FROM videos WHERE id=?", (video_id,)).fetchone()
-            if vrow and vrow['people_json']:
-                plist = json.loads(vrow['people_json'])
-                new_plist = [rename_map.get(x, x) for x in plist]
-                if new_plist != plist:
+                "SELECT people_json, summary_json FROM videos WHERE id=?",
+                (video_id,)).fetchone()
+            if vrow:
+                v_changed = False
+                new_people_json = vrow['people_json']
+                if vrow['people_json']:
+                    plist = json.loads(vrow['people_json'])
+                    new_plist = [rename_map.get(x, x) for x in plist]
+                    if new_plist != plist:
+                        new_people_json = json.dumps(new_plist, ensure_ascii=False)
+                        v_changed = True
+                new_summary = _rewrite_text(vrow['summary_json'])
+                if new_summary != vrow['summary_json']:
+                    v_changed = True
+                if v_changed:
                    self._conn.execute(
-                        "UPDATE videos SET people_json=?, updated_at=? WHERE id=?",
-                        (json.dumps(new_plist, ensure_ascii=False), now, video_id))
+                        "UPDATE videos SET people_json=?, summary_json=?, updated_at=? "
+                        "WHERE id=?",
+                        (new_people_json, new_summary, now, video_id))
            self._conn.commit()

    def correct_video_identity(self, video_id: int, current_name: str, new_name: str):
--- a/fam-edge/tests/test_oracle_db.py
+++ b/fam-edge/tests/test_oracle_db.py
@@ -267,6 +267,58 @@ def test_rewrite_event_person_names_updates_events_and_video(tmp_path):
    assert set(json.loads(vrow["people_json"])) == {"爷爷", "媳妇"}


+def test_rewrite_event_person_names_rewrites_description_text(tmp_path):
+    """核心诉求: description 是大模型写的自然语言句子，"人物A/人物B"这类 uid
+    会直接以文字形式嵌在句子里，只改 person_list_json/person_appearances_json
+    这些结构化字段的话，事件卡片上方徽章显示对了，描述文字里还是旧 uid，两处
+    对不上——description 也要做文本替换。"""
+    db = _db(tmp_path)
+    vid = db.ensure_video("motion_2_2000.mp4", "/tmp/motion_2_2000.mp4",
+                          event_start_time="2026-08-22 13:00:00")
+    events = [
+        {"timestamp": "13:29:24",
+         "description": "人物B双手叉腰站在客厅中央；人物A在远处厨房；儿童已离开画面。",
+         "people": ["人物A", "人物B"],
+         "person_appearances": [
+             {"uid": "人物A", "features": {"gender": "男"}, "action": "站立"},
+             {"uid": "人物B", "features": {"gender": "女"}, "action": "叉腰"}]},
+    ]
+    db.mark_video_processed(vid, "人物A和人物B都在客厅活动。", events,
+                            ["人物A", "人物B"], "gemini")
+    db.rewrite_event_person_names(vid, {"人物A": "爸爸", "人物B": "媳妇"})
+
+    ev_row = db._conn.execute(
+        "SELECT description FROM events WHERE video_id=?", (vid,)).fetchone()
+    assert ev_row["description"] == "媳妇双手叉腰站在客厅中央；爸爸在远处厨房；儿童已离开画面。"
+
+    v_row = db._conn.execute(
+        "SELECT summary_json FROM videos WHERE id=?", (vid,)).fetchone()
+    assert v_row["summary_json"] == "爸爸和媳妇都在客厅活动。"
+
+
+def test_rewrite_event_person_names_longer_labels_replaced_before_shorter(tmp_path):
+    """核心诉求: uid 可能带 "#2"/"#3" 这类同名冲突后缀，"人物A" 是 "人物A#2" 的
+    前缀——如果先替换短的 "人物A"，"人物A#2" 会被错误地部分命中变成"爷爷#2"，
+    而不是走它自己在 rename_map 里对应的正确目标。必须长的先替换。"""
+    db = _db(tmp_path)
+    vid = db.ensure_video("motion_3_3000.mp4", "/tmp/motion_3_3000.mp4",
+                          event_start_time="2026-08-22 13:00:00")
+    events = [
+        {"timestamp": "13:00:01",
+         "description": "人物A和人物A#2一起在客厅。",
+         "people": ["人物A", "人物A#2"],
+         "person_appearances": [
+             {"uid": "人物A", "features": {"gender": "男"}, "action": "站立"},
+             {"uid": "人物A#2", "features": {"gender": "女"}, "action": "站立"}]},
+    ]
+    db.mark_video_processed(vid, "摘要", events, ["人物A", "人物A#2"], "gemini")
+    db.rewrite_event_person_names(vid, {"人物A": "爷爷", "人物A#2": "媳妇"})
+
+    ev_row = db._conn.execute(
+        "SELECT description FROM events WHERE video_id=?", (vid,)).fetchone()
+    assert ev_row["description"] == "爷爷和媳妇一起在客厅。"
+
+
 def test_rewrite_event_person_names_noop_on_empty_map(tmp_path):
    db = _db(tmp_path)
    vid = _seed_video_with_events(db)