feat: 添加报告压缩功能解决上下文超限问题

ReportEngine/agent.py

@@ -234,13 +234,20 @@ def __init__(self, config: Optional[Settings] = None):
 
         # 初始化节点
         self._initialize_nodes()
-        
+
         # 初始化文件数量基准
         self._initialize_file_baseline()
-
+
+        # 初始化报告压缩器
+        from .utils.report_compressor import ReportCompressor
+        self.report_compressor = ReportCompressor(
+            config=self.config,
+            llm_client=self.llm_client if self.config.SUMMARY_STRATEGY == "llm" else None
+        )
+
         # 状态
         self.state = ReportState()
-        
+
         # GraphRAG 状态数据（每次 load_input_files 时重置）
         self._loaded_states = {}
 
@@ -476,6 +483,14 @@ def generate_report(
 
         normalized_reports = self._normalize_reports(reports)
 
+        # 根据配置决定是否启用报告压缩
+        if self.config.ENABLE_REPORT_COMPRESSION:
+            # 为阶段1-2生成摘要版本
+            summarized_reports = self.report_compressor.summarize_reports(normalized_reports)
+            logger.info("已生成报告摘要版本用于文档设计和篇幅规划")
+        else:
+            summarized_reports = normalized_reports
+
         def emit(event_type: str, payload: Dict[str, Any]):
             """面向Report Engine流通道的事件分发器，保证错误不外泄。"""
             if not stream_handler:
@@ -516,7 +531,7 @@ def emit(event_type: str, payload: Dict[str, Any]):
                 lambda: self.document_layout_node.run(
                     sections,
                     template_text,
-                    normalized_reports,
+                    summarized_reports,  # 使用摘要版本
                     forum_logs,
                     query,
                     template_overview,
@@ -536,7 +551,7 @@ def emit(event_type: str, payload: Dict[str, Any]):
                 lambda: self.word_budget_node.run(
                     sections,
                     layout_design,
-                    normalized_reports,
+                    summarized_reports,  # 使用摘要版本
                     forum_logs,
                     query,
                     template_overview,
@@ -558,7 +573,7 @@ def emit(event_type: str, payload: Dict[str, Any]):
 
             generation_context = self._build_generation_context(
                 query,
-                normalized_reports,
+                normalized_reports,  # 传递原始报告，在章节生成时动态提取
                 forum_logs,
                 template_result,
                 layout_design,
@@ -1071,6 +1086,8 @@ def _build_generation_context(
             "query": query,
             "template_name": template_result.get("template_name"),
             "reports": reports,
+            "reports_original": reports,  # 保留原始报告用于章节生成时的动态提取
+            "report_compressor": self.report_compressor,  # 传递压缩器实例
             "forum_logs": self._stringify(forum_logs),
             "theme_tokens": theme_tokens,
             "style_directives": {

ReportEngine/nodes/chapter_generation_node.py

@@ -306,7 +306,30 @@ def _build_payload(self, section: TemplateSection, context: Dict[str, Any]) -> D
         返回:
             dict: 可以直接序列化进提示词的payload，兼顾章节信息与全局约束。
         """
+        # 获取报告内容
+        reports_original = context.get("reports_original", {})
         reports = context.get("reports", {})
+
+        # 如果启用压缩且有原始报告，则提取相关内容
+        compressor = context.get("report_compressor")
+        if compressor and reports_original:
+            try:
+                # 从配置中获取压缩开关
+                from ..utils.config import settings
+                if settings.ENABLE_REPORT_COMPRESSION:
+                    # 动态提取与章节相关的内容
+                    reports = compressor.extract_relevant_content(
+                        reports=reports_original,
+                        chapter_title=section.title,
+                        chapter_outline=section.outline
+                    )
+                    logger.info(
+                        f"章节 '{section.title}' 已提取相关内容"
+                    )
+            except Exception as e:
+                logger.warning(f"章节内容提取失败: {e}，使用原始报告")
+                reports = reports_original
+
         # 章节篇幅规划（来自WordBudgetNode），用于指导字数与强调点
         chapter_plan_map = context.get("chapter_directives", {})
         chapter_plan = chapter_plan_map.get(section.chapter_id) if chapter_plan_map else {}
@@ -355,7 +378,7 @@ def _build_payload(self, section: TemplateSection, context: Dict[str, Any]) -> D
             "chapterPlan": chapter_plan,
             "wordPlan": context.get("word_plan"),
         }
-        
+
         # GraphRAG 增强：如果上下文中包含图谱查询结果，添加到payload
         graph_results = context.get("graph_results")
         if graph_results:
@@ -370,7 +393,7 @@ def _build_payload(self, section: TemplateSection, context: Dict[str, Any]) -> D
             graph_enhancement = context.get("graph_enhancement_prompt")
             if graph_enhancement:
                 payload["graphEnhancementPrompt"] = graph_enhancement
-        
+
         if chapter_plan:
             constraints = payload["constraints"]
             if chapter_plan.get("targetWords"):

ReportEngine/utils/config.py

@@ -75,6 +75,32 @@ class Settings(BaseSettings):
         default=3, description="GraphRAG每章节查询次数上限"
     )
 
+    # 报告压缩配置
+    ENABLE_REPORT_COMPRESSION: bool = Field(
+        default=True, description="是否启用报告压缩以避免上下文超限"
+    )
+    SUMMARY_STRATEGY: str = Field(
+        default="rule", description="摘要策略：rule=规则提取, llm=LLM摘要, hybrid=混合"
+    )
+    SUMMARY_COMPRESSION_RATIO: float = Field(
+        default=0.35, description="摘要目标压缩率（0.3-0.4推荐）"
+    )
+    EXTRACTION_STRATEGY: str = Field(
+        default="keyword", description="提取策略：keyword=关键词匹配, embedding=语义相似度"
+    )
+    EXTRACTION_MAX_RATIO: float = Field(
+        default=0.5, description="章节提取内容最大比例（相对原文）"
+    )
+    KEYWORD_MATCH_THRESHOLD: int = Field(
+        default=2, description="段落至少匹配的关键词数量"
+    )
+    KEEP_CONTEXT_PARAGRAPHS: bool = Field(
+        default=True, description="提取时是否保留匹配段落的上下文"
+    )
+    CONTEXT_PARAGRAPHS_COUNT: int = Field(
+        default=1, description="保留的上下文段落数量（前后各N段）"
+    )
+
     class Config:
         """Pydantic配置：允许从.env读取并兼容大小写"""
         env_file = ".env"
@@ -109,5 +135,10 @@ def print_config(config: Settings):
     message += f"PDF 导出: {config.ENABLE_PDF_EXPORT}\n"
     message += f"图表样式: {config.CHART_STYLE}\n"
     message += f"LLM API Key: {'已配置' if config.REPORT_ENGINE_API_KEY else '未配置'}\n"
+    message += f"报告压缩: {config.ENABLE_REPORT_COMPRESSION}\n"
+    if config.ENABLE_REPORT_COMPRESSION:
+        message += f"  摘要策略: {config.SUMMARY_STRATEGY}\n"
+        message += f"  压缩率: {config.SUMMARY_COMPRESSION_RATIO:.0%}\n"
+        message += f"  提取策略: {config.EXTRACTION_STRATEGY}\n"
     message += "=========================\n"
     logger.info(message)