重构斗鱼粉丝日报信息提纯链路

- 新增本地弹幕文件测试入口，支持直接对样本文件生成提纯结果
- 将本地统计、主题证据簇和语义事实提示接入斗鱼日报LLM材料
- 明确降低情绪刷屏权重，改为优先提取赛事、位置、英雄、对局和场外互动信息

plugins/douyu/danmu_summary.py

@@ -24,6 +24,49 @@ class DouyuDanmuSummaryHelper:
         "666", "6666", "牛", "牛逼", "稳", "寄", "杀", "帅", "好", "行", "绷", "哭", "乐",
         "哈哈", "哈哈哈", "笑死", "卧槽", "wc", "awsl", "nb", "nbl", "c", "6",
     }
+    # 这些词对“直播间气氛”有价值，但对“事实提炼”帮助有限。
+    # 后面在内容线索排序时会适当降权，避免把真正的赛事/英雄/剧情信息淹掉。
+    GENERIC_REACTION_TERMS: Set[str] = {
+        "哈哈", "哈哈哈", "哈哈哈哈", "哈哈哈哈哈", "哈哈哈哈哈哈",
+        "gg", "g", "888", "1", "啊", "啊？", "坏了", "好起来了", "翻了",
+    }
+    # 高频语义簇配置：
+    # 1. 不做中文分词，而是直接按“直播圈常见话题簇”收证据；
+    # 2. 每个簇都会保留计数、时间范围和原声样本；
+    # 3. 这样模型更容易抓到“今天到底发生了哪些具体事”，而不是只看到大量情绪词。
+    FACT_CLUSTER_CONFIGS: List[Dict[str, Any]] = [
+        {
+            "label": "赛事预告与报名动态",
+            "keywords": ["老头杯", "选人", "报名", "开赛", "比赛", "30号", "4月30", "4月30日"],
+        },
+        {
+            "label": "比赛位置与身份讨论",
+            "keywords": ["1号位", "5号位", "打1", "打5", "教练", "carry", "辅助"],
+        },
+        {
+            "label": "镜头与外形调侃",
+            "keywords": ["摄像头", "开摄像头", "光头", "秃头", "洗头", "面容", "露脸"],
+        },
+        {
+            "label": "团播人物与场外关系",
+            "keywords": ["糯糯", "瑶瑶", "冬瓜", "冬瓜强", "白队", "团播", "户外"],
+        },
+        {
+            "label": "关键对局与局势转折",
+            "keywords": ["奶绿", "muerta", "gg", "翻了", "拿下", "上高地", "守高地", "队友", "大炮", "萨尔", "炸弹人"],
+        },
+    ]
+    HERO_ALIASES: Dict[str, List[str]] = {
+        "Muerta/奶绿": ["奶绿", "muerta"],
+        "德鲁伊/Lone Druid": ["德鲁伊", "lone druid", "熊德"],
+        "小小/Tiny": ["小小", "tiny"],
+        "帕克/Puck": ["帕克", "puck"],
+        "火猫/Ember Spirit": ["火猫", "ember"],
+        "敌法/Anti-Mage": ["敌法", "am", "anti-mage"],
+        "兽王/Beastmaster": ["兽王", "beastmaster"],
+        "萨尔/Disruptor": ["萨尔", "disruptor"],
+        "炸弹人/Techies": ["炸弹人", "techies"],
+    }
     NOISE_PATTERNS = [
         re.compile(r"本条弹幕.*机器人", re.I),
         re.compile(r"请不要.*统计机器人数", re.I),
@@ -110,6 +153,26 @@ class DouyuDanmuSummaryHelper:
                     collected.append(parsed)
         return collected
 
+    @classmethod
+    def load_messages_from_file(cls, file_path: str) -> List[Dict[str, Any]]:
+        """
+        从指定文本文件直接读取弹幕。
+        这个入口主要用于本地调试和样本回归：
+        1. 不依赖 Redis session；
+        2. 不要求文件落在 temp/douyu_danmu 目录；
+        3. 便于直接拿用户提供的测试样本跑提纯链路。
+        """
+        path = str(file_path or "").strip()
+        if not path or not os.path.exists(path):
+            return []
+        collected: List[Dict[str, Any]] = []
+        with open(path, "r", encoding="utf-8", errors="ignore") as f:
+            for line in f:
+                parsed = cls.parse_danmu_line(line)
+                if parsed:
+                    collected.append(parsed)
+        return collected
+
     @classmethod
     def build_summary_material(
         cls,
@@ -289,6 +352,8 @@ class DouyuDanmuSummaryHelper:
                 burst_terms=burst_terms,
                 limit=cue_limit,
             ),
+            # 再补一层“事实型提示”，专门抬高赛事、位置、英雄、镜头梗、关键对局等信息密度高的内容。
+            "semantic_fact_hints": cls._build_semantic_fact_hints(ordered_messages),
         }
 
     @staticmethod
@@ -1199,7 +1264,7 @@ class DouyuDanmuSummaryHelper:
 
         for item in repeated_messages:
             push(
-                "repeat",
+                "emotion" if cls._normalize_template_text(str(item.get("text") or "")) in cls.GENERIC_REACTION_TERMS else "repeat",
                 str(item.get("text") or ""),
                 int(item.get("count", 0) or 0),
                 int(item.get("user_count", 0) or 0),
@@ -1222,7 +1287,7 @@ class DouyuDanmuSummaryHelper:
             if count < 2:
                 continue
             push(
-                "short_repeat",
+                "emotion" if normalized in cls.GENERIC_REACTION_TERMS else "short_repeat",
                 short_message_text_map.get(normalized, normalized),
                 count,
                 len([uid for uid in short_message_users.get(normalized, set()) if uid]),
@@ -1230,15 +1295,142 @@ class DouyuDanmuSummaryHelper:
 
         for item in burst_terms:
             push(
-                "burst",
+                "emotion" if cls._normalize_template_text(str(item.get("text") or "")) in cls.GENERIC_REACTION_TERMS else "burst",
                 str(item.get("text") or ""),
                 int(item.get("count", 0) or 0),
                 int(item.get("user_count", 0) or 0),
             )
 
-        cues.sort(key=lambda item: (int(item.get("count", 0) or 0), int(item.get("user_count", 0) or 0)), reverse=True)
+        kind_priority = {
+            "repeat": 5,
+            "short_repeat": 4,
+            "burst": 3,
+            "emotion": 1,
+        }
+        cues.sort(
+            key=lambda item: (
+                int(kind_priority.get(str(item.get("kind") or ""), 0)),
+                int(item.get("count", 0) or 0),
+                int(item.get("user_count", 0) or 0),
+            ),
+            reverse=True,
+        )
         return cues[:limit]
 
+    @classmethod
+    def _build_semantic_fact_hints(cls, messages: List[Dict[str, Any]]) -> Dict[str, Any]:
+        """
+        生成“事实型语义提示”。
+        这层不是做总结，而是把模型容易漏掉的高价值信息提前挂出来：
+        1. 赛事/选人/位置讨论；
+        2. 英雄与关键局；
+        3. 摄像头、团播人物等场外互动梗。
+        """
+        return {
+            "topic_clusters": cls._build_fact_topic_clusters(messages),
+            "hero_mentions": cls._build_hero_mentions(messages),
+        }
+
+    @classmethod
+    def _build_fact_topic_clusters(cls, messages: List[Dict[str, Any]], limit: int = 8) -> List[Dict[str, Any]]:
+        clusters: List[Dict[str, Any]] = []
+        for config in cls.FACT_CLUSTER_CONFIGS:
+            matched_items: List[Dict[str, Any]] = []
+            keywords = [str(item).lower() for item in (config.get("keywords") or []) if str(item).strip()]
+            for item in messages:
+                content = str(item.get("content") or "").strip()
+                if not content:
+                    continue
+                lowered = content.lower()
+                if any(keyword in lowered for keyword in keywords):
+                    matched_items.append(item)
+            if not matched_items:
+                continue
+            sample_messages = []
+            seen = set()
+            for item in matched_items:
+                content = str(item.get("content") or "").strip()
+                normalized = cls._normalize_template_text(content)
+                if not normalized or normalized in seen:
+                    continue
+                seen.add(normalized)
+                sample_messages.append({
+                    "date": cls._format_date(item.get("timestamp")),
+                    "hm": cls._format_hm(item.get("timestamp")),
+                    "nickname": str(item.get("nickname") or "").strip(),
+                    "content": content[:100],
+                })
+                if len(sample_messages) >= 5:
+                    break
+            if not sample_messages:
+                continue
+            clusters.append({
+                "label": str(config.get("label") or "").strip(),
+                "match_count": len(matched_items),
+                "user_count": len({
+                    str(item.get("uid") or "") for item in matched_items if str(item.get("uid") or "").strip()
+                }),
+                "first_hm": cls._format_hm(matched_items[0].get("timestamp")),
+                "last_hm": cls._format_hm(matched_items[-1].get("timestamp")),
+                "keywords": config.get("keywords", [])[:8],
+                "samples": sample_messages,
+            })
+        clusters.sort(
+            key=lambda item: (
+                int(item.get("match_count", 0) or 0),
+                int(item.get("user_count", 0) or 0),
+            ),
+            reverse=True,
+        )
+        return clusters[:limit]
+
+    @classmethod
+    def _build_hero_mentions(cls, messages: List[Dict[str, Any]], limit: int = 6) -> List[Dict[str, Any]]:
+        hero_results: List[Dict[str, Any]] = []
+        for hero_name, aliases in cls.HERO_ALIASES.items():
+            matched_items: List[Dict[str, Any]] = []
+            alias_list = [str(alias).lower() for alias in aliases if str(alias).strip()]
+            for item in messages:
+                content = str(item.get("content") or "").strip()
+                if not content:
+                    continue
+                lowered = content.lower()
+                if any(alias in lowered for alias in alias_list):
+                    matched_items.append(item)
+            if not matched_items:
+                continue
+            samples = []
+            seen = set()
+            for item in matched_items:
+                content = str(item.get("content") or "").strip()
+                normalized = cls._normalize_template_text(content)
+                if not normalized or normalized in seen:
+                    continue
+                seen.add(normalized)
+                samples.append({
+                    "hm": cls._format_hm(item.get("timestamp")),
+                    "nickname": str(item.get("nickname") or "").strip(),
+                    "content": content[:100],
+                })
+                if len(samples) >= 4:
+                    break
+            hero_results.append({
+                "hero": hero_name,
+                "mention_count": len(matched_items),
+                "user_count": len({
+                    str(item.get("uid") or "") for item in matched_items if str(item.get("uid") or "").strip()
+                }),
+                "samples": samples,
+            })
+        hero_results.sort(
+            key=lambda item: (
+                int(item.get("mention_count", 0) or 0),
+                int(item.get("user_count", 0) or 0),
+            ),
+            reverse=True,
+        )
+        return hero_results[:limit]
+
     @staticmethod
     def _looks_like_pure_punctuation(content: str) -> bool:
         text = str(content or "").strip()