优化斗鱼弹幕总结：新增粉丝向弹幕萃取区块并调整提示词语气

plugins/douyu/report_template.py

@@ -24,19 +24,37 @@ def _render_list(items: List[str], item_class: str = "bullet-list") -> str:
     return f'<ul class="{item_class}">{lis}</ul>' if lis else ""
 
 
-def _split_summary_blocks(danmu_summary: str) -> tuple[str, List[str]]:
+def _split_summary_blocks(danmu_summary: str) -> tuple[str, List[str], List[str]]:
+    # 这里把 LLM 返回的弹幕总结拆成三部分：
+    # 1) lead: 顶部总述段落
+    # 2) insight_items: 常规的复盘要点（运营/观察视角）
+    # 3) fans_extract_items: 专门给粉丝看的“弹幕萃取”要点
+    # 约定：当检测到“【粉丝向弹幕萃取】”或同义标记后，后续条目归入 fans_extract_items。
     lead_parts = []
     insight_items = []
+    fans_extract_items = []
+    in_fans_extract_block = False
     for line in str(danmu_summary or "").splitlines():
         stripped = line.strip()
         if not stripped:
             continue
+        # 兼容不同模型可能产出的标题样式，尽量把粉丝向内容稳定识别出来。
+        if stripped.startswith("【粉丝向弹幕萃取】") or stripped.startswith("粉丝向弹幕萃取") or stripped.startswith("给粉丝看的弹幕萃取"):
+            in_fans_extract_block = True
+            continue
         if stripped.startswith("- "):
-            insight_items.append(stripped[2:].strip())
+            if in_fans_extract_block:
+                fans_extract_items.append(stripped[2:].strip())
+            else:
+                insight_items.append(stripped[2:].strip())
         else:
-            lead_parts.append(stripped)
+            # 非 bullet 文本在粉丝区块中也保留，避免模型偶发输出短段落导致信息丢失。
+            if in_fans_extract_block:
+                fans_extract_items.append(stripped)
+            else:
+                lead_parts.append(stripped)
     lead = " ".join(lead_parts).strip()
-    return lead, insight_items
+    return lead, insight_items, fans_extract_items
 
 
 def _normalize_summary_bullets(payload: Dict[str, Any], items: List[str], target_count: int = 5) -> List[str]:
@@ -75,6 +93,47 @@ def _normalize_summary_bullets(payload: Dict[str, Any], items: List[str], target
     return normalized[:target_count]
 
 
+def _normalize_fans_extract_bullets(payload: Dict[str, Any], items: List[str], target_count: int = 6) -> List[str]:
+    # 粉丝向萃取强调“现场感”，优先保留模型给出的条目；
+    # 不足时再从代表弹幕/重复梗中补齐，避免页面出现空区块。
+    normalized = [str(item or "").strip() for item in items if str(item or "").strip()]
+    if len(normalized) >= target_count:
+        return normalized[:target_count]
+
+    supplements: List[str] = []
+    representative_messages = payload.get("representative_messages", []) or []
+    repeated_messages = payload.get("repeated_messages", []) or []
+    burst_terms = payload.get("burst_terms", []) or []
+
+    for item in representative_messages[:8]:
+        nickname = str(item.get("nickname") or "").strip() or "观众"
+        content = str(item.get("content") or "").strip()
+        if not content:
+            continue
+        supplements.append(f"{nickname}：{content[:46]}")
+
+    for item in repeated_messages[:6]:
+        text = str(item.get("text") or "").strip()
+        count = int(item.get("count", 0) or 0)
+        if text:
+            supplements.append(f"复读梗「{text[:34]}」出现 {count} 次。")
+
+    for item in burst_terms[:4]:
+        text = str(item.get("text") or "").strip()
+        count = int(item.get("count", 0) or 0)
+        if text:
+            supplements.append(f"情绪短词「{text}」集中刷了 {count} 次。")
+
+    existing = set(normalized)
+    for item in supplements:
+        if item not in existing:
+            normalized.append(item)
+            existing.add(item)
+        if len(normalized) >= target_count:
+            break
+    return normalized[:target_count]
+
+
 def _build_template_items(payload: Dict[str, Any], limit: int = 8) -> List[str]:
     items: List[str] = []
     seen = set()
@@ -400,8 +459,9 @@ def render_daily_report_html(
     top_active_users = payload.get("operator_metrics", {}).get("top_active_users", []) or []
     audience_trend = payload.get("audience_trend", {}) or {}
 
-    lead_summary, danmu_bullets = _split_summary_blocks(danmu_summary)
+    lead_summary, danmu_bullets, fans_extract_bullets = _split_summary_blocks(danmu_summary)
     danmu_bullets = _normalize_summary_bullets(payload, danmu_bullets, target_count=5)
+    fans_extract_bullets = _normalize_fans_extract_bullets(payload, fans_extract_bullets, target_count=6)
 
     html_doc = f"""<html>
 <head>
@@ -589,6 +649,30 @@ def render_daily_report_html(
         grid-template-columns: repeat(2, minmax(0, 1fr));
         gap: 12px;
     }}
+    .fans-panel {{
+        margin-top: 14px;
+        padding: 14px 15px 12px;
+        border-radius: 18px;
+        background: linear-gradient(180deg, rgba(255,255,255,0.96), rgba(245,250,255,0.94));
+        border: 1px solid rgba(73, 136, 224, 0.18);
+    }}
+    .fans-title {{
+        color: #1d4ed8;
+        font-size: 13px;
+        letter-spacing: .06em;
+        font-weight: 700;
+        margin-bottom: 8px;
+    }}
+    .fans-list {{
+        margin: 0;
+        padding-left: 18px;
+    }}
+    .fans-list li {{
+        color: #1e3a5f;
+        margin: 8px 0;
+        line-height: 1.65;
+        font-size: 14px;
+    }}
     .insight-card {{
         padding: 15px 16px;
         border-radius: 18px;
@@ -946,6 +1030,10 @@ def render_daily_report_html(
                         <div class="insight-grid">
                             {_render_insight_cards(danmu_bullets)}
                         </div>
+                        <div class="fans-panel">
+                            <div class="fans-title">给粉丝看的弹幕萃取</div>
+                            {_render_list(fans_extract_bullets, item_class="fans-list")}
+                        </div>
                     </div>
                     <div class="aside-card">
                         <div class="aside-title">高频梗</div>