修复采集器流程顺序: 采集→分析→LLM总结

- _generate_topic_with_llm 改名为 _generate_topics_with_llm
- 支持同时接收已分析的案例+搜索结果作为LLM上下文
- run()顺序改为: 采集→RSS提炼案例→LLM基于全部数据生成→降级回退
- LLM现在在流程末尾做总结生成,而非开头替代搜索

scripts/collector.py

@@ -377,13 +377,13 @@ class SustainabilityCollector:
             logger.warning(f"web_search失败 {source.name}: {e}")
             return []
 
-    def _generate_topic_with_llm(self, search_results: Optional[List[Dict]] = None) -> Optional[SustainabilityTopic]:
+    def _generate_topics_with_llm(self, cases: List[SustainabilityCase] = None, search_results: List[Dict] = None) -> List[SustainabilityTopic]:
-        """用LLM生成选题（有搜索结果时参考，无结果时直接生成）"""
+        """用LLM基于采集数据生成选题（数据充分时精确生成，无数据时凭知识生成）"""
         try:
             from app.core.nvidia_client import call_llm
         except ImportError:
             logger.warning("LLM不可用，跳过AI选题生成")
-            return None
+            return []
 
         existing = self._get_existing_titles()
         existing_hint = ""
@@ -394,30 +394,28 @@ class SustainabilityCollector:
         day_idx = datetime.datetime.now().timetuple().tm_yday % len(categories)
         target_category = categories[day_idx]
 
-        search_section = ""
+        data_section = ""
         if search_results:
-            summaries = [f"- {r.get('title','')}: {r.get('content','')[:120]}" for r in search_results[:4]]
+            summaries = [f"- {r.get('title','')}: {r.get('content','')[:100]}" for r in search_results[:5]]
-            search_section = "搜索结果参考：\n" + "\n".join(summaries) + "\n"
+            data_section += "搜索结果：\n" + "\n".join(summaries) + "\n"
+        if cases:
+            case_lines = [f"- {c.title[:40]}({c.category})" for c in cases[:5]]
+            data_section += "\n采集案例：\n" + "\n".join(case_lines) + "\n"
 
-        prompt = f"""你是一个内容策略师。生成一个面向中国年轻读者、有价值、适合传播的选题。
+        prompt = f"""你是一个内容策略师。基于以下信息，为「{target_category}」类别生成一个高质量选题。
 
-目标类别：{target_category}
+{data_section if data_section else "（当前无实时采集数据，请基于你对中文互联网趋势的了解直接生成）"}
-{search_section}
 {existing_hint}
 
-生成一个选题，直接输出JSON（不要其他文字）：
+输出一个选题，格式JSON：
 {{{{
-  "title": "标题（20字内，含核心关键词，避免「新趋势」「指南」这类烂尾词）",
+  "title": "标题（20字内，含核心关键词）",
-  "core_concept": "核心观点（一句话说清独特价值）",
+  "core_concept": "核心观点（一句话）",
-  "audience_pain": "受众痛点（真实用户的困惑）",
+  "audience_pain": "受众痛点",
   "unique_angle": "差异化切入点",
   "format": "内容形式（趋势洞察/实操指南/对比分析/案例解读）"
 }}}}
-
+只输出JSON。"""
-要求：
-- 标题像普通人会搜索的
-- 切入点具体，不泛泛而谈
-- 优先考虑中国读者能实操的内容"""
 
         try:
             resp = call_llm(prompt, temperature=0.7)
@@ -431,7 +429,7 @@ class SustainabilityCollector:
             topic = SustainabilityTopic(
                 id=topic_id,
                 title=data.get("title", f"{target_category}新观察"),
-                cases=[],
+                cases=[c.id for c in (cases or [])[:3]],
                 audience="城市焦虑青年(26-35岁)",
                 china_pain_points=data.get("audience_pain", ""),
                 localization_solution="文章中将提供具体可执行的建议",
@@ -456,10 +454,10 @@ class SustainabilityCollector:
                 platform_urls={}
             )
             logger.info(f"LLM生成选题: {topic.title}")
-            return topic
+            return [topic]
         except Exception as e:
             logger.warning(f"LLM选题生成失败: {e}")
-            return None
+            return []
 
     def analyze_article(self, article: Dict) -> Optional[SustainabilityCase]:
         """分析文章内容,提炼案例"""
@@ -786,24 +784,24 @@ class SustainabilityCollector:
         logger.info(f"RSS采集 {sum(1 for a in all_articles if a.get('source_name','') not in [s.name for s in self.sources if s.type=='web_search'])} 篇, "
                     f"搜索采集 {len(web_search_results)} 篇")
 
-        # ---------------------- 第二阶段：LLM选题生成 ----------------------
+        # ---------------------- 第二阶段：RSS文章提炼案例 ----------------------
-        llm_topic = self._generate_topic_with_llm(web_search_results if web_search_results else None)
-
-        if llm_topic and not self._is_duplicate_topic(llm_topic.title, existing_titles):
-            llm_topic.created_at = datetime.datetime.now().isoformat()
-            llm_topic.lock_by = None
-            llm_topic.lock_at = None
-            llm_topic.status = "待处理"
-            self.new_topics.append(llm_topic)
-            logger.info(f"✅ LLM生成选题: {llm_topic.title}")
-
-        # ---------------------- 第三阶段：RSS文章提炼案例 ----------------------
         rss_articles = [a for a in all_articles if a not in web_search_results]
         for article in rss_articles[:15]:
             case = self.analyze_article(article)
             if case:
                 self.new_cases.append(case)
 
+        # ---------------------- 第三阶段：LLM基于采集数据生成选题 ----------------------
+        llm_topics = self._generate_topics_with_llm(cases=self.new_cases, search_results=web_search_results)
+        for topic in llm_topics:
+            if not self._is_duplicate_topic(topic.title, existing_titles):
+                topic.created_at = datetime.datetime.now().isoformat()
+                topic.lock_by = None
+                topic.lock_at = None
+                topic.status = "待处理"
+                self.new_topics.append(topic)
+                logger.info(f"✅ LLM生成选题: {topic.title}")
+
         # ---------------------- 第四阶段：降级策略 ----------------------
         if not self.new_topics and len(self.new_cases) < 2:
             logger.warning(f"LLM选题和RSS案例均不足,启动本地案例降级")