backend 去ai化

backend/app/services/matching_service.py

@@ -1,7 +1,3 @@
-"""
-匹配服务：根据用户兴趣关键词（精确 + 语义）推荐事件
-打分融合：标签/标题匹配分 + 标签相关度 + 热度 + 新鲜度加成
-"""
 import os
 from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
@@ -13,6 +9,7 @@ from sqlalchemy.orm import Session
 from app.models.models import ExtractedTopic, TargetType, UnifiedEvent, UserTopicPreference, utcnow
 from app.services.fetcher_service import embedder_model
 
+# AI辅助生成：deepseek-v3-2，2026年3月20日
 
 # 语义匹配阈值：用户关键词和事件标签/标题向量相似度达到该值才计入语义命中
 DEFAULT_PREFERENCE_SEMANTIC_THRESHOLD = 0.78
@@ -35,6 +32,7 @@ class MatchedEventResult:
     semantic_hits: list[dict[str, Any]]
     tags: list[str]
 
+# AI生成结束
 
 def _normalize_text(text: str) -> str:
     """统一小写与首尾空白，便于做稳定匹配。"""
@@ -80,7 +78,6 @@ def _build_keyword_embedding_map(keywords: list[str]) -> dict[str, np.ndarray]:
         
     uncached_keywords = []
     
-    # 1. 尝试从缓存获取
     for keyword in keywords:
         if not keyword:
             continue
@@ -89,9 +86,7 @@ def _build_keyword_embedding_map(keywords: list[str]) -> dict[str, np.ndarray]:
         else:
             uncached_keywords.append(keyword)
             
-    # 2. 对未命中的词进行统一的批量推理
     if uncached_keywords:
-        # 去重，避免同一个未缓存的词被计算多次
         unique_uncached = list(dict.fromkeys(uncached_keywords))
         
         vectors = embedder_model.encode(unique_uncached, normalize_embeddings=True, show_progress_bar=False)
@@ -102,7 +97,6 @@ def _build_keyword_embedding_map(keywords: list[str]) -> dict[str, np.ndarray]:
             for k in keys_to_delete:
                 del _EMBEDDING_CACHE[k]
                 
-        # 3. 将新计算的向量存入缓存并回填结果
         for keyword, vec in zip(unique_uncached, vectors):
             vec_array = np.asarray(vec, dtype=np.float32)
             _EMBEDDING_CACHE[keyword] = vec_array
@@ -172,7 +166,6 @@ def recommend_events_for_user(
         else PREFERENCE_SEMANTIC_THRESHOLD
     )
 
-    # 1. 读取用户兴趣词
     preferences = (
         db.query(UserTopicPreference)
         .filter(UserTopicPreference.user_id == user_id)
@@ -185,7 +178,6 @@ def recommend_events_for_user(
     if not preference_keywords:
         return []
 
-    # 2. 读取候选事件（时间 + 热度过滤，避免全表扫描）
     time_limit = utcnow() - timedelta(hours=hours)
     events = (
         db.query(UnifiedEvent)
@@ -213,20 +205,17 @@ def recommend_events_for_user(
         .all()
     )
 
-    # 组织事件标签映射：event_id -> [(tag, relevance_score), ...]
     event_topics: dict[int, list[tuple[str, float | None]]] = {}
     for event_id, topic_keyword, relevance_score in topic_rows:
         if not topic_keyword:
             continue
         event_topics.setdefault(event_id, []).append((topic_keyword, relevance_score))
 
-    # 3. 批量编码用户词与标签词，减少模型调用次数
     unique_preference_keywords = list(dict.fromkeys(preference_keywords))
     unique_topic_keywords = list(dict.fromkeys([row[1] for row in topic_rows if row[1]]))
     pref_vec_map = _build_keyword_embedding_map(unique_preference_keywords)
     topic_vec_map = _build_keyword_embedding_map(unique_topic_keywords)
 
-    # 预先建立“标准化后用户词集合”，用于精确匹配
     normalized_preference_pairs = [
         (word, _normalize_text(word))
         for word in unique_preference_keywords
@@ -246,20 +235,15 @@ def recommend_events_for_user(
         exact_hits: list[str] = []
         semantic_hits: list[dict[str, Any]] = []
         score = 0.0
-
-        # 对每个事件标签做精确匹配或语义匹配
         for topic_keyword, topic_relevance in topic_list:
             topic_relevance_score = float(topic_relevance) if topic_relevance is not None else 50.0
-
-            # 1) 精确命中（包括完全相等与包含关系）
+            
             matched_pref = _find_exact_preference_match(topic_keyword, normalized_preference_pairs)
             if matched_pref is not None:
                 exact_hits.append(topic_keyword)
-                # 精确命中给较高基础分，标签自身相关度作为增益
                 score += 45.0 + topic_relevance_score * 0.2
                 continue
 
-            # 2) 语义命中（未精确命中时再算）
             best_pref, best_sim = _find_best_semantic_match(topic_keyword, topic_vec_map, pref_vec_map)
 
             if best_pref is not None and best_sim >= similarity_threshold:
@@ -270,10 +254,8 @@ def recommend_events_for_user(
                         "similarity": round(best_sim, 4),
                     }
                 )
-                # 语义命中分略低于精确命中，并由相似度放大
                 score += best_sim * 35.0 + topic_relevance_score * 0.12
 
-        # 标题也参与匹配，但权重低于结构化标签，避免长标题过度主导排序。
         event_title = (event.unified_title or "").strip()
         if event_title:
             title_exact_pref = _find_exact_preference_match(event_title, normalized_preference_pairs)
@@ -292,15 +274,12 @@ def recommend_events_for_user(
                     )
                     score += best_sim * 24.0
 
-        # 如果精确和语义都没命中，直接跳过
         if not exact_hits and not semantic_hits:
             continue
 
-        # 融合事件热度和新鲜度，避免只看语义分
         score += min(event.hot_score, 100) * 0.3
         score += _calc_freshness_bonus(event)
 
-        # 返回标签时做去重，保证接口稳定
         tags = list(dict.fromkeys([item[0] for item in topic_list]))
         scored_results.append(
             MatchedEventResult(