From 2a36e4ec3d600fc94e26c0db89a511264fdf8651 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=B0=8F=E9=B1=BC=E5=BC=80=E5=8F=91?= <git@xiaoyu.com>
Date: Sun, 17 May 2026 21:35:44 +0800
Subject: [PATCH] =?UTF-8?q?fix(material):=20=E6=94=AF=E6=8C=81=20scene=20?=
 =?UTF-8?q?=E5=90=8D=E7=A7=B0=E9=A1=BA=E5=BA=8F=E9=A2=A0=E5=80=92=E5=85=9C?=
 =?UTF-8?q?=E5=BA=95=E5=8C=B9=E9=85=8D?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

AI 生成 scene 时常将三级分类名中的 '-' 前后顺序写反
（如 瓷砖铺贴-瓷砖完工展示 vs 瓷砖完工展示-瓷砖铺贴），
导致精确匹配失败、素材匹配为空。

- match_material: 精确匹配失败后，尝试倒序匹配
- batch_match: 批量查询时同时查询原始名和倒序名，
  内存中构建 scene->category 映射，优先精确匹配、fallback 倒序
---
 python-api/app/services/material_service.py | 50 ++++++++++++++++++---
 1 file changed, 43 insertions(+), 7 deletions(-)

diff --git a/python-api/app/services/material_service.py b/python-api/app/services/material_service.py
index c2cccea..441d196 100644
--- a/python-api/app/services/material_service.py
+++ b/python-api/app/services/material_service.py
@@ -94,10 +94,22 @@ async def match_material(
 
     normalized = _normalize_scene(scene)
 
-    # 1. 查找三级分类
+    # 1. 查找三级分类（精确匹配 + 顺序颠倒兜底）
     category = await broll_category.get_by_name_and_level(
         db, name=normalized, level=3
     )
+    # 若精确匹配失败，尝试将 "A-B" 倒序为 "B-A" 再匹配
+    if category is None:
+        parts = normalized.rsplit("-", 1)
+        if len(parts) == 2:
+            reversed_name = f"{parts[1]}-{parts[0]}"
+            category = await broll_category.get_by_name_and_level(
+                db, name=reversed_name, level=3
+            )
+            if category:
+                logger.info(
+                    f"素材分类顺序颠倒兜底命中: '{normalized}' -> '{reversed_name}'"
+                )
     if category is None:
         logger.debug(f"未找到分类: {normalized}")
         return None
@@ -175,17 +187,41 @@ async def batch_match(
     normalized_scenes = [_normalize_scene(s["scene"]) for s in scenes]
     unique_names = list(set(normalized_scenes))
 
-    # 2. 批量查询分类（1 次 DB）
+    # 2. 批量查询分类（1 次 DB）—— 同时查询原始名和倒序名
+    reversed_names: list[str] = []
+    name_to_reversed: dict[str, str] = {}
+    for name in unique_names:
+        parts = name.rsplit("-", 1)
+        if len(parts) == 2:
+            rev = f"{parts[1]}-{parts[0]}"
+            reversed_names.append(rev)
+            name_to_reversed[name] = rev
+
+    all_query_names = unique_names + reversed_names
     categories = await broll_category.get_by_names_and_level(
-        db, names=unique_names, level=3
+        db, names=all_query_names, level=3
     )
-    category_map = {c.name: c for c in categories}
+    category_map: dict[str, object] = {}
+    for c in categories:
+        category_map[c.name] = c
+
+    # 构建原始 scene -> category 的映射（优先精确匹配，fallback 倒序匹配）
+    scene_to_category: dict[str, object] = {}
+    for name in unique_names:
+        if name in category_map:
+            scene_to_category[name] = category_map[name]
+        elif name in name_to_reversed and name_to_reversed[name] in category_map:
+            rev = name_to_reversed[name]
+            scene_to_category[name] = category_map[rev]
+            logger.info(
+                f"批量匹配顺序颠倒兜底命中: '{name}' -> '{rev}'"
+            )
 
     # 3. 收集所有需要的 category_id
     needed_category_ids = [
-        category_map[name].id
+        scene_to_category[name].id
         for name in unique_names
-        if name in category_map
+        if name in scene_to_category
     ]
 
     # 4. 批量查询素材（1 次 DB）
@@ -215,7 +251,7 @@ async def batch_match(
     for idx, scene_name in enumerate(normalized_scenes):
         required_duration = scenes[idx]["duration"]
 
-        category = category_map.get(scene_name)
+        category = scene_to_category.get(scene_name)
         if category is None:
             results.append(None)
             continue