From 4386d0df75348b6ee6f601ee387731d378f5dfb3 Mon Sep 17 00:00:00 2001 From: liuwei Date: Wed, 29 Apr 2026 15:06:56 +0800 Subject: [PATCH] =?UTF-8?q?=E9=87=8D=E6=9E=84=E6=96=97=E9=B1=BC=E7=B2=89?= =?UTF-8?q?=E4=B8=9D=E6=97=A5=E6=8A=A5=E4=B8=BA=E4=BF=A1=E6=81=AF=E4=BC=98?= =?UTF-8?q?=E5=85=88=E7=BB=93=E6=9E=84?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 1. 更新粉丝日报提示词,优先提炼赛事、位置、英雄、对局和场外有效信息\n2. 扩展模板解析与渲染逻辑,支持今日重点信息、核心讨论话题、英雄与对局焦点等新板块\n3. 优化粉丝日报兜底文案与模板展示,让本地提纯结果和LLM语义总结共同参与输出 --- plugins/douyu/main.py | 69 ++++++++-- plugins/douyu/report_template.py | 123 +++++++++++++++++- .../douyu/templates/daily_fans_report.html | 14 ++ 3 files changed, 193 insertions(+), 13 deletions(-) diff --git a/plugins/douyu/main.py b/plugins/douyu/main.py index 2aa007e..bb6f71a 100644 --- a/plugins/douyu/main.py +++ b/plugins/douyu/main.py @@ -2306,32 +2306,36 @@ class DouyuPlugin(MessagePluginInterface): """ 粉丝版日报提示词设计目标: 1. 和运营版彻底区分开,不再强调“策略、复盘、活跃质量”; - 2. 保留真实弹幕语境,让输出像“群友拿着回放在整活”; + 2. 先提炼高价值信息,再保留粉丝向乐子感,避免报告只剩几条段子; 3. 允许轻微恶搞和夸张,但不能编造未出现的事件,也不能攻击主播或观众。 """ meta = payload.get("report_meta", {}) or {} room_context_prompt = self._build_room_context_prompt_block(payload) prompt_material = self._build_llm_prompt_material(payload, include_operator=False) system_prompt = ( - "你是斗鱼直播间的粉丝向整活日报编辑。" - "请只根据提供的真实弹幕材料,输出一份开心、欢乐、带一点恶搞气质的中文总结。" - "语气要像群友在复盘名场面,不要写成运营分析,不要编造剧情,不要使用代码块。" + "你是斗鱼直播间的粉丝向信息日报编辑。" + "请只根据提供的真实弹幕材料,输出一份既有信息量、又保留直播间欢乐气氛的中文总结。" + "语气要像群友在复盘直播名场面,但第一优先级是提炼有效信息,不要写成运营分析,不要编造剧情,不要使用代码块。" "如果这是 Dota2 / 电竞语境直播间,请优先按刀圈/电竞圈人物关系、职业生涯、老比赛和主播互动梗去理解笑点。" ) user_prompt = ( - "请输出一份适合给粉丝看的《斗鱼弹幕乐子日报》,严格按下面结构输出:\n" + "请输出一份适合给粉丝看的《斗鱼弹幕信息日报》,严格按下面结构输出:\n" "1. 开头先写 1 段总述,概括今天直播间的整体节目效果和气氛。\n" - "2. 另起一行写标题:`【今日笑点】`,下面写 4 条 bullet,每条一句,突出最有节目效果的地方。\n" - "3. 另起一行写标题:`【弹幕名场面】`,下面写 4-6 条 bullet,尽量保留弹幕原话风格,像现场回放。\n" - "4. 另起一行写标题:`【梗王榜】`,下面写 3 条 bullet,把今天最刷屏、最有共识的梗排出来。\n" - "5. 另起一行写标题:`【收尾播报】`,下面只写 1 句收尾,轻松一点,像群里发图后的总结句。\n" - "6. 可以夸张一点、调皮一点,但不要低俗,不要攻击主播,不要使用“建议、策略、转化、数据表现”等运营词。\n\n" + "2. 另起一行写标题:`【今日重点信息】`,下面写 4-6 条 bullet,优先提炼真正有效的信息。重点看赛事预告、具体日期、位置讨论、人物关系、主播近况、是否开摄像头、场外话题等。\n" + "3. 另起一行写标题:`【核心讨论话题】`,下面写 3-4 条 bullet,概括今天弹幕主要围绕哪些话题打转,每条都要带具体内容,不要空泛。\n" + "4. 另起一行写标题:`【英雄与对局焦点】`,下面写 3-4 条 bullet,提炼今天重点英雄、关键对局走势、翻盘/崩盘点、观众对操作和出装的主要反馈。\n" + "5. 另起一行写标题:`【今日笑点】`,下面写 3-4 条 bullet,每条一句,突出最有节目效果的地方。\n" + "6. 另起一行写标题:`【弹幕名场面】`,下面写 4-6 条 bullet,尽量保留弹幕原话风格,像现场回放。\n" + "7. 另起一行写标题:`【梗王榜】`,下面写 3 条 bullet,把今天最刷屏、最有共识的梗排出来。\n" + "8. 另起一行写标题:`【收尾播报】`,下面只写 1 句收尾,轻松一点,像群里发图后的总结句。\n" + "9. 出现时间信息时,尽量写清楚绝对日期或明确时间,比如“4月30日”“18:45 前后”,不要只写“最近”“那天”。\n" + "10. 不要写“建议、策略、转化、数据表现”等运营词,也不要只复述哈哈哈、gg 这种已经能由本地统计完成的噪声。\n\n" f"主播:{meta.get('nickname') or meta.get('room_name') or meta.get('room_id')}\n" f"日期:{meta.get('anchor_day', '')}\n" f"{room_context_prompt}" "下面是已经提纯给 LLM 的现场材料,请优先抓 `topic_evidence_clusters` 和 `compact_scene_material` 里的 `semantic_fact_hints`、原声弹幕、时间线块和集体起哄片段," "尤其留意赛事预告、位置讨论、英雄选择、关键对局、镜头调侃和团播人物关系," - "少写空泛概括。\n" + "少写空泛概括。若材料无法支持某个判断,就不要写。\n" f"材料:\n{json.dumps(prompt_material, ensure_ascii=False, indent=2)}" ) return system_prompt, user_prompt @@ -2824,6 +2828,13 @@ class DouyuPlugin(MessagePluginInterface): 兜底文本保持“有梗但不胡编”的原则,所有句子都只从真实弹幕统计结果里取材。 """ meta = payload.get("report_meta", {}) or {} + topic_clusters = payload.get("topic_evidence_clusters", []) or [] + hero_mentions = ( + payload.get("compact_scene_material", {}) + .get("semantic_fact_hints", {}) + .get("hero_mentions", []) + or [] + ) top_terms = [ str(item.get("term") or "").strip() for item in (payload.get("top_terms", []) or [])[:5] @@ -2844,8 +2855,42 @@ class DouyuPlugin(MessagePluginInterface): f"尤其是「{str(merged_templates[0].get('text') or '').strip()[:26]}」这类共识弹幕,一看就是全场默认会背。" ) - lines = [" ".join(lead_parts).strip(), "【今日笑点】"] + lines = [" ".join(lead_parts).strip(), "【今日重点信息】"] + for item in topic_clusters[:3]: + label = str(item.get("label") or "").strip() + time_range = str(item.get("time_range") or "").strip() + count = int(item.get("count", 0) or 0) + samples = item.get("samples", []) or [] + sample_text = "" + if samples: + sample_text = str(samples[0].get("content") or "").strip()[:38] + if label and sample_text: + lines.append(f"- {label}从 {time_range or '全场'} 一直有人聊,约 {count} 条相关弹幕,代表说法是「{sample_text}」。") + elif label: + lines.append(f"- {label}是今天的重点主线之一,相关弹幕约 {count} 条。") + + lines.append("【核心讨论话题】") + for item in topic_clusters[:3]: + label = str(item.get("label") or "").strip() + keywords = [str(keyword).strip() for keyword in (item.get("keywords", []) or [])[:5] if str(keyword).strip()] + if label and keywords: + lines.append(f"- 大家围着 {label} 打转,关键词主要是 {'、'.join(keywords)}。") + + lines.append("【英雄与对局焦点】") + for item in hero_mentions[:3]: + hero_name = str(item.get("hero") or "").strip() + mention_count = int(item.get("mention_count", 0) or 0) + samples = item.get("samples", []) or [] + sample_text = "" + if samples: + sample_text = str(samples[0].get("content") or "").strip()[:36] + if hero_name and sample_text: + lines.append(f"- {hero_name}被点名 {mention_count} 次,弹幕现场直接说到「{sample_text}」。") + elif hero_name: + lines.append(f"- {hero_name}是今天的主要英雄话题之一,被提到 {mention_count} 次。") + + lines.append("【今日笑点】") if peak_buckets: top_bucket = peak_buckets[0] lines.append( diff --git a/plugins/douyu/report_template.py b/plugins/douyu/report_template.py index d3597e2..14a5154 100644 --- a/plugins/douyu/report_template.py +++ b/plugins/douyu/report_template.py @@ -91,6 +91,15 @@ def _split_fans_report_blocks(report_text: str) -> Dict[str, Any]: 即便模型没有完全按约定输出,这里也会尽量兜底,保证页面不空。 """ header_alias_map = { + "今日重点信息": "key_info", + "重点信息": "key_info", + "有效信息": "key_info", + "核心讨论话题": "topic_focus", + "讨论话题": "topic_focus", + "核心话题": "topic_focus", + "英雄与对局焦点": "hero_focus", + "对局焦点": "hero_focus", + "英雄焦点": "hero_focus", "今日笑点": "laugh_points", "笑点": "laugh_points", "欢乐总结": "laugh_points", @@ -106,6 +115,9 @@ def _split_fans_report_blocks(report_text: str) -> Dict[str, Any]: } sections = { "lead": "", + "key_info": [], + "topic_focus": [], + "hero_focus": [], "laugh_points": [], "famous_scenes": [], "meme_rank": [], @@ -510,6 +522,97 @@ def _build_fans_effective_info_lines(payload: Dict[str, Any], limit: int = 6) -> return lines[:limit] +def _build_local_topic_focus_lines(payload: Dict[str, Any], limit: int = 4) -> List[str]: + """ + 为“核心讨论话题”补充本地可直接确定的摘要句。 + 这里故意不让模型自己重新发明事实,而是把主题簇已经聚好的结果转成人能读懂的话。 + """ + lines: List[str] = [] + seen = set() + + def push(text: str) -> None: + value = str(text or "").strip() + if not value or value in seen: + return + seen.add(value) + lines.append(value) + + for item in (payload.get("topic_evidence_clusters", []) or [])[:4]: + label = str(item.get("label") or "").strip() + keywords = [str(keyword).strip() for keyword in (item.get("keywords", []) or [])[:5] if str(keyword).strip()] + count = int(item.get("count", 0) or 0) + if label and keywords: + push(f"{label}是高频主线,相关讨论约 {count} 条,关键词集中在 {'、'.join(keywords)}。") + elif label: + push(f"{label}是今天反复被拉出来聊的主线之一,相关讨论约 {count} 条。") + if len(lines) >= limit: + return lines[:limit] + return lines[:limit] + + +def _build_local_hero_focus_lines(payload: Dict[str, Any], limit: int = 4) -> List[str]: + """ + 为“英雄与对局焦点”准备本地兜底。 + 这部分直接复用英雄提及聚类,优先强调出现频次和代表发言,方便粉丝快速看懂今天在聊什么英雄。 + """ + hero_mentions = ( + payload.get("compact_scene_material", {}) + .get("semantic_fact_hints", {}) + .get("hero_mentions", []) + or [] + ) + lines: List[str] = [] + seen = set() + + def push(text: str) -> None: + value = str(text or "").strip() + if not value or value in seen: + return + seen.add(value) + lines.append(value) + + for item in hero_mentions[:4]: + hero_name = str(item.get("hero") or "").strip() + mention_count = int(item.get("mention_count", 0) or 0) + samples = item.get("samples", []) or [] + sample_text = "" + if samples: + sample_text = str(samples[0].get("content") or "").strip()[:36] + if hero_name and sample_text: + push(f"{hero_name}被提到 {mention_count} 次,现场典型弹幕是「{sample_text}」。") + elif hero_name: + push(f"{hero_name}是今天的主要英雄讨论点之一,被提到 {mention_count} 次。") + if len(lines) >= limit: + return lines[:limit] + return lines[:limit] + + +def _normalize_information_section_items( + llm_items: List[str], + local_items: List[str], + target_count: int, +) -> List[str]: + """ + 将模型提炼结果与本地事实兜底合并。 + 设计目标: + 1. 先尊重模型已经总结好的“可读句子”; + 2. 如果模型漏了,就用本地证据补足; + 3. 始终保证最终区块有信息量,而不是空标题。 + """ + normalized: List[str] = [] + seen = set() + for source in (llm_items, local_items): + for item in source: + value = str(item or "").strip() + if not value or value in seen: + continue + seen.add(value) + normalized.append(value) + if len(normalized) >= target_count: + return normalized[:target_count] + return normalized[:target_count] + + def _render_fans_info_cards(items: List[str]) -> str: blocks = [] for item in items[:6]: @@ -929,6 +1032,21 @@ def render_fans_daily_report_html( f" | 围观群众 {meta.get('unique_user_count', 0)} 人" ) sections = _split_fans_report_blocks(fans_report_text) + effective_info_lines = _normalize_information_section_items( + sections.get("key_info", []), + _build_fans_effective_info_lines(payload), + target_count=6, + ) + topic_focus_lines = _normalize_information_section_items( + sections.get("topic_focus", []), + _build_local_topic_focus_lines(payload), + target_count=4, + ) + hero_focus_lines = _normalize_information_section_items( + sections.get("hero_focus", []), + _build_local_hero_focus_lines(payload), + target_count=4, + ) laugh_points = _normalize_funny_bullets(payload, sections.get("laugh_points", []), target_count=4) famous_scenes = _normalize_scene_bullets(payload, sections.get("famous_scenes", []), target_count=5) meme_rank = _normalize_rank_bullets(payload, sections.get("meme_rank", []), target_count=3) @@ -954,8 +1072,11 @@ def render_fans_daily_report_html( "lead_text": lead_text, # 粉丝版不再只做“乐子文案展示”,而是补进本地提纯后的有效信息区。 "fans_metrics_html": Markup(_render_fans_metric_cards(_build_fans_fun_metrics(payload))), - "effective_info_html": Markup(_render_fans_info_cards(_build_fans_effective_info_lines(payload))), + "effective_summary_html": Markup(_render_list(effective_info_lines, item_class="section-summary-list")), + "effective_info_html": Markup(_render_fans_info_cards(effective_info_lines)), + "topic_focus_html": Markup(_render_list(topic_focus_lines, item_class="section-summary-list")), "topic_clusters_html": Markup(_render_topic_clusters(topic_clusters)), + "hero_focus_html": Markup(_render_list(hero_focus_lines, item_class="section-summary-list")), "hero_mentions_html": Markup(_render_hero_mentions(hero_mentions)), "hot_windows_html": Markup(_render_hot_window_cards(local_stats.get("peak_windows", []) or [])), "repeat_digest_html": Markup(_render_repeat_digest(payload)), diff --git a/plugins/douyu/templates/daily_fans_report.html b/plugins/douyu/templates/daily_fans_report.html index 536aa69..10ccb74 100644 --- a/plugins/douyu/templates/daily_fans_report.html +++ b/plugins/douyu/templates/daily_fans_report.html @@ -139,6 +139,17 @@ background: linear-gradient(180deg, rgba(255,255,255,.96), rgba(255,249,244,.94)); border: 1px solid var(--line); } + .section-summary-list { + margin: 0 0 14px; + padding-left: 22px; + } + .section-summary-list li { + margin: 8px 0; + color: #5a3e37; + font-size: 15px; + line-height: 1.76; + font-weight: 600; + } .section-title { display: flex; align-items: center; @@ -387,16 +398,19 @@
今日重点信息
+ {{ effective_summary_html }}
{{ effective_info_html }}
核心讨论话题
+ {{ topic_focus_html }}
{{ topic_clusters_html }}
英雄与对局焦点
+ {{ hero_focus_html }}
{{ hero_mentions_html }}