feat：添加lpmm内部接口，信息抽取类和一个测试脚本

scripts/lpmm_interactive_manager.py

@@ -0,0 +1,265 @@
+import asyncio
+import os
+import sys
+
+# 尽量统一控制台编码为 utf-8，避免中文输出报错
+try:
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+    if hasattr(sys.stderr, "reconfigure"):
+        sys.stderr.reconfigure(encoding="utf-8")
+except Exception:
+    pass
+
+# 确保项目根目录在 sys.path 中，以便导入 src.*
+PROJECT_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
+if PROJECT_ROOT not in sys.path:
+    sys.path.append(PROJECT_ROOT)
+
+try:
+    # 显式从 src.chat.knowledge.lpmm_ops 导入单例对象
+    from src.chat.knowledge.lpmm_ops import lpmm_ops
+    from src.common.logger import get_logger
+    from src.memory_system.retrieval_tools.query_lpmm_knowledge import query_lpmm_knowledge
+    from src.chat.knowledge import lpmm_start_up
+    from src.config.config import global_config
+except ImportError as e:
+    print(f"导入失败，请确保在项目根目录下运行脚本: {e}")
+    sys.exit(1)
+
+logger = get_logger("lpmm_interactive_manager")
+
+async def interactive_add():
+    """交互式导入知识"""
+    print("\n" + "=" * 40)
+    print("      --- 📥 导入知识 (Add) ---")
+    print("=" * 40)
+    print("说明：请输入要导入的文本内容。")
+    print("      - 支持多段落，段落间请保留空行。")
+    print("      - 输入完成后，在新起的一行输入 'EOF' 并回车结束输入。")
+    print("-" * 40)
+    
+    lines = []
+    while True:
+        try:
+            line = input()
+            if line.strip().upper() == "EOF":
+                break
+            lines.append(line)
+        except EOFError:
+            break
+            
+    text = "\n".join(lines).strip()
+    if not text:
+        print("\n[!] 内容为空，操作已取消。")
+        return
+
+    print("\n[进度] 正在调用 LPMM 接口进行信息抽取与向量化，请稍候...")
+    try:
+        # 使用 lpmm_ops.py 中的接口
+        result = await lpmm_ops.add_content(text)
+        
+        if result["status"] == "success":
+            print(f"\n[√] 成功：{result['message']}")
+            print(f"    实际新增段落数: {result.get('count', 0)}")
+        else:
+            print(f"\n[×] 失败：{result['message']}")
+    except Exception as e:
+        print(f"\n[×] 发生异常: {e}")
+        logger.error(f"add_content 异常: {e}", exc_info=True)
+
+async def interactive_delete():
+    """交互式删除知识"""
+    print("\n" + "=" * 40)
+    print("      --- 🗑️ 删除知识 (Delete) ---")
+    print("=" * 40)
+    print("删除模式：")
+    print("  1. 关键词模糊匹配（删除包含关键词的所有段落）")
+    print("  2. 完整文段匹配（删除完全匹配的段落）")
+    print("-" * 40)
+    
+    mode = input("请选择删除模式 (1/2): ").strip()
+    exact_match = False
+    
+    if mode == "2":
+        exact_match = True
+        print("\n[完整文段匹配模式]")
+        print("说明：请输入要删除的完整文段内容（必须完全一致）。")
+        print("      - 支持多行输入，输入完成后在新起的一行输入 'EOF' 并回车。")
+        print("-" * 40)
+        lines = []
+        while True:
+            try:
+                line = input()
+                if line.strip().upper() == "EOF":
+                    break
+                lines.append(line)
+            except EOFError:
+                break
+        keyword = "\n".join(lines).strip()
+    else:
+        if mode != "1":
+            print("\n[!] 无效选择，默认使用关键词模糊匹配模式。")
+        print("\n[关键词模糊匹配模式]")
+        keyword = input("请输入匹配关键词: ").strip()
+    
+    if not keyword:
+        print("\n[!] 输入为空，操作已取消。")
+        return
+        
+    print("-" * 40)
+    confirm = input(f"危险确认：确定要删除所有匹配 '{keyword[:50]}{'...' if len(keyword) > 50 else ''}' 的知识吗？(y/N): ").strip().lower()
+    if confirm != 'y':
+        print("\n[!] 已取消删除操作。")
+        return
+
+    print("\n[进度] 正在执行删除并更新索引...")
+    try:
+        # 使用 lpmm_ops.py 中的接口
+        result = await lpmm_ops.delete(keyword, exact_match=exact_match)
+        
+        if result["status"] == "success":
+            print(f"\n[√] 成功：{result['message']}")
+            print(f"    删除条数: {result.get('deleted_count', 0)}")
+        elif result["status"] == "info":
+            print(f"\n[i] 提示：{result['message']}")
+        else:
+            print(f"\n[×] 失败：{result['message']}")
+    except Exception as e:
+        print(f"\n[×] 发生异常: {e}")
+        logger.error(f"delete 异常: {e}", exc_info=True)
+
+async def interactive_clear():
+    """交互式清空知识库"""
+    print("\n" + "=" * 40)
+    print("      --- ⚠️ 清空知识库 (Clear All) ---")
+    print("=" * 40)
+    print("警告：此操作将删除LPMM知识库中的所有内容！")
+    print("      - 所有段落向量")
+    print("      - 所有实体向量")
+    print("      - 所有关系向量")
+    print("      - 整个知识图谱")
+    print("      - 此操作不可恢复！")
+    print("-" * 40)
+    
+    # 双重确认
+    confirm1 = input("⚠️  第一次确认：确定要清空整个知识库吗？(输入 'YES' 继续): ").strip()
+    if confirm1 != "YES":
+        print("\n[!] 已取消清空操作。")
+        return
+    
+    print("\n" + "=" * 40)
+    confirm2 = input("⚠️  第二次确认：此操作不可恢复，请再次输入 'CLEAR' 确认: ").strip()
+    if confirm2 != "CLEAR":
+        print("\n[!] 已取消清空操作。")
+        return
+    
+    print("\n[进度] 正在清空知识库...")
+    try:
+        # 使用 lpmm_ops.py 中的接口
+        result = await lpmm_ops.clear_all()
+        
+        if result["status"] == "success":
+            print(f"\n[√] 成功：{result['message']}")
+            stats = result.get("stats", {})
+            before = stats.get("before", {})
+            after = stats.get("after", {})
+            print("\n[统计信息]")
+            print(f"  清空前: 段落={before.get('paragraphs', 0)}, 实体={before.get('entities', 0)}, "
+                  f"关系={before.get('relations', 0)}, KG节点={before.get('kg_nodes', 0)}, KG边={before.get('kg_edges', 0)}")
+            print(f"  清空后: 段落={after.get('paragraphs', 0)}, 实体={after.get('entities', 0)}, "
+                  f"关系={after.get('relations', 0)}, KG节点={after.get('kg_nodes', 0)}, KG边={after.get('kg_edges', 0)}")
+        else:
+            print(f"\n[×] 失败：{result['message']}")
+    except Exception as e:
+        print(f"\n[×] 发生异常: {e}")
+        logger.error(f"clear_all 异常: {e}", exc_info=True)
+
+async def interactive_search():
+    """交互式查询知识"""
+    print("\n" + "=" * 40)
+    print("      --- 🔍 查询知识 (Search) ---")
+    print("=" * 40)
+    print("说明：输入查询问题或关键词，系统会返回相关的知识段落。")
+    print("-" * 40)
+    
+    # 确保 LPMM 已初始化
+    if not global_config.lpmm_knowledge.enable:
+        print("\n[!] 警告：LPMM 知识库在配置中未启用。")
+        return
+    
+    try:
+        lpmm_start_up()
+    except Exception as e:
+        print(f"\n[!] LPMM 初始化失败: {e}")
+        logger.error(f"LPMM 初始化失败: {e}", exc_info=True)
+        return
+    
+    query = input("请输入查询问题或关键词: ").strip()
+    
+    if not query:
+        print("\n[!] 查询内容为空，操作已取消。")
+        return
+    
+    # 询问返回条数
+    print("-" * 40)
+    limit_str = input("希望返回的相关知识条数（默认3，直接回车使用默认值）: ").strip()
+    try:
+        limit = int(limit_str) if limit_str else 3
+        limit = max(1, min(limit, 20))  # 限制在1-20之间
+    except ValueError:
+        limit = 3
+        print("[!] 输入无效，使用默认值 3。")
+    
+    print("\n[进度] 正在查询知识库...")
+    try:
+        result = await query_lpmm_knowledge(query, limit=limit)
+        
+        print("\n" + "=" * 60)
+        print("[查询结果]")
+        print("=" * 60)
+        print(result)
+        print("=" * 60)
+    except Exception as e:
+        print(f"\n[×] 查询失败: {e}")
+        logger.error(f"查询异常: {e}", exc_info=True)
+
+async def main():
+    """主循环"""
+    while True:
+        print("\n" + "╔" + "═" * 38 + "╗")
+        print("║      LPMM 知识库交互管理工具        ║")
+        print("╠" + "═" * 38 + "╣")
+        print("║  1. 导入知识 (Add Content)          ║")
+        print("║  2. 删除知识 (Delete Content)       ║")
+        print("║  3. 查询知识 (Search Content)       ║")
+        print("║  4. 清空知识库 (Clear All) ⚠️        ║")
+        print("║  0. 退出 (Exit)                     ║")
+        print("╚" + "═" * 38 + "╝")
+        
+        choice = input("请选择操作编号: ").strip()
+        
+        if choice == "1":
+            await interactive_add()
+        elif choice == "2":
+            await interactive_delete()
+        elif choice == "3":
+            await interactive_search()
+        elif choice == "4":
+            await interactive_clear()
+        elif choice in ("0", "q", "Q", "quit", "exit"):
+            print("\n已退出工具。")
+            break
+        else:
+            print("\n[!] 无效的选择，请输入 0, 1, 2, 3 或 4。")
+
+if __name__ == "__main__":
+    try:
+        # 运行主循环
+        asyncio.run(main())
+    except KeyboardInterrupt:
+        print("\n\n[!] 用户中断程序 (Ctrl+C)。")
+    except Exception as e:
+        print(f"\n[!] 程序运行出错: {e}")
+        logger.error(f"Main loop 异常: {e}", exc_info=True)
+

src/chat/knowledge/ie_process.py

@@ -1,7 +1,7 @@
 import asyncio
 import json
 import time
-from typing import List, Union
+from typing import List, Union, Dict, Any
 
 from .global_logger import logger
 from . import prompt_template
@@ -173,3 +173,50 @@ def info_extract_from_str(
                 return None, None
 
     return entity_extract_result, rdf_triple_extract_result
+
+
+class IEProcess:
+    """
+    信息抽取处理器类，提供更方便的批次处理接口。
+    """
+
+    def __init__(self, llm_ner: LLMRequest, llm_rdf: LLMRequest = None):
+        self.llm_ner = llm_ner
+        self.llm_rdf = llm_rdf or llm_ner
+
+    async def process_paragraphs(self, paragraphs: List[str]) -> List[dict]:
+        """
+        异步处理多个段落。
+        """
+        from .utils.hash import get_sha256
+
+        results = []
+        total = len(paragraphs)
+        
+        for i, pg in enumerate(paragraphs, start=1):
+            # 打印进度日志，让用户知道没有卡死
+            logger.info(f"[IEProcess] 正在处理第 {i}/{total} 段文本 (长度: {len(pg)})...")
+            
+            # 使用 asyncio.to_thread 包装同步阻塞调用，防止死锁
+            # 这样 info_extract_from_str 内部的 asyncio.run 会在独立线程的新 loop 中运行
+            try:
+                entities, triples = await asyncio.to_thread(
+                    info_extract_from_str, self.llm_ner, self.llm_rdf, pg
+                )
+
+                if entities is not None:
+                    results.append(
+                        {
+                            "idx": get_sha256(pg),
+                            "passage": pg,
+                            "extracted_entities": entities,
+                            "extracted_triples": triples,
+                        }
+                    )
+                    logger.info(f"[IEProcess] 第 {i}/{total} 段处理完成，提取到 {len(entities)} 个实体")
+                else:
+                    logger.warning(f"[IEProcess] 第 {i}/{total} 段提取失败（返回为空）")
+            except Exception as e:
+                logger.error(f"[IEProcess] 处理第 {i}/{total} 段时发生异常: {e}")
+
+        return results

src/chat/knowledge/lpmm_ops.py

@@ -0,0 +1,331 @@
+import asyncio
+from typing import List, Callable, Any
+from src.chat.knowledge.embedding_store import EmbeddingManager
+from src.chat.knowledge.kg_manager import KGManager
+from src.chat.knowledge.qa_manager import QAManager
+from src.common.logger import get_logger
+from src.config.config import global_config
+from src.chat.knowledge import get_qa_manager, lpmm_start_up
+
+logger = get_logger("LPMM-Plugin-API")
+
+class LPMMOperations:
+    """
+    LPMM 内部操作接口。
+    封装了 LPMM 的核心操作，供插件系统 API 或其他内部组件调用。
+    """
+    
+    def __init__(self):
+        self._initialized = False
+
+    async def _run_cancellable_executor(
+        self, func: Callable, *args, **kwargs
+    ) -> Any:
+        """
+        在线程池中执行可取消的同步操作。
+        当任务被取消时（如 Ctrl+C），会立即响应并抛出 CancelledError。
+        注意：线程池中的操作可能仍在运行，但协程会立即返回，不会阻塞主进程。
+        
+        Args:
+            func: 要执行的同步函数
+            *args: 函数的位置参数
+            **kwargs: 函数的关键字参数
+            
+        Returns:
+            函数的返回值
+            
+        Raises:
+            asyncio.CancelledError: 当任务被取消时
+        """
+        loop = asyncio.get_event_loop()
+        # 在线程池中执行，当协程被取消时会立即响应
+        # 虽然线程池中的操作可能仍在运行，但协程不会阻塞
+        return await loop.run_in_executor(None, func, *args, **kwargs)
+
+    async def _get_managers(self) -> tuple[EmbeddingManager, KGManager, QAManager]:
+        """获取并确保 LPMM 管理器已初始化"""
+        qa_mgr = get_qa_manager()
+        if qa_mgr is None:
+            # 如果全局没初始化，尝试初始化
+            if not global_config.lpmm_knowledge.enable:
+                logger.warning("LPMM 知识库在全局配置中未启用，操作可能受限。")
+            
+            lpmm_start_up()
+            qa_mgr = get_qa_manager()
+            
+        if qa_mgr is None:
+            raise RuntimeError("无法获取 LPMM QAManager，请检查 LPMM 是否已正确安装和配置。")
+            
+        return qa_mgr.embed_manager, qa_mgr.kg_manager, qa_mgr
+
+    async def add_content(self, text: str) -> dict:
+        """
+        向知识库添加新内容。
+        
+        Args:
+            text: 原始文本。支持多段文本（用双换行分隔）。
+            
+        Returns:
+            dict: {"status": "success/error", "count": 导入段落数, "message": "描述"}
+        """
+        try:
+            embed_mgr, kg_mgr, _ = await self._get_managers()
+            
+            # 1. 分段处理
+            paragraphs = [p.strip() for p in text.split('\n\n') if p.strip()]
+            if not paragraphs:
+                return {"status": "error", "message": "文本内容为空"}
+
+            # 2. 实体与三元组抽取 (内部调用大模型)
+            from src.chat.knowledge.ie_process import IEProcess
+            from src.llm_models.utils_model import LLMRequest
+            from src.config.config import model_config
+            
+            llm_ner = LLMRequest(model_set=model_config.model_task_config.lpmm_entity_extract, request_type="lpmm.entity_extract")
+            llm_rdf = LLMRequest(model_set=model_config.model_task_config.lpmm_rdf_build, request_type="lpmm.rdf_build")
+            ie_process = IEProcess(llm_ner, llm_rdf)
+            
+            logger.info(f"[Plugin API] 正在对 {len(paragraphs)} 段文本执行信息抽取...")
+            extracted_docs = await ie_process.process_paragraphs(paragraphs)
+            
+            # 3. 构造并导入数据
+            # 这里我们手动实现导入逻辑，不依赖外部脚本
+            # a. 准备段落
+            raw_paragraphs = {doc["idx"]: doc["passage"] for doc in extracted_docs}
+            # b. 准备三元组
+            triple_list_data = {doc["idx"]: doc["extracted_triples"] for doc in extracted_docs}
+
+            # 向量化并入库
+            # 注意：此处模仿 import_openie.py 的核心逻辑
+            # 1. 先进行去重检查，只处理新段落
+            # store_new_data_set 期望的格式：raw_paragraphs 的键是段落hash（不带前缀），值是段落文本
+            new_raw_paragraphs = {}
+            new_triple_list_data = {}
+            
+            for pg_hash, passage in raw_paragraphs.items():
+                key = f"paragraph-{pg_hash}"
+                if key not in embed_mgr.stored_pg_hashes:
+                    new_raw_paragraphs[pg_hash] = passage
+                    new_triple_list_data[pg_hash] = triple_list_data[pg_hash]
+
+            if not new_raw_paragraphs:
+                return {"status": "success", "count": 0, "message": "内容已存在，无需重复导入"}
+
+            # 2. 使用 EmbeddingManager 的标准方法存储段落、实体和关系的嵌入
+            # store_new_data_set 会自动处理嵌入生成和存储
+            # 将同步阻塞操作放到线程池中执行，避免阻塞事件循环
+            await self._run_cancellable_executor(
+                embed_mgr.store_new_data_set,
+                new_raw_paragraphs,
+                new_triple_list_data
+            )
+
+            # 3. 构建知识图谱（只需要三元组数据和embedding_manager）
+            await self._run_cancellable_executor(
+                kg_mgr.build_kg,
+                new_triple_list_data,
+                embed_mgr
+            )
+
+            # 4. 持久化
+            await self._run_cancellable_executor(embed_mgr.rebuild_faiss_index)
+            await self._run_cancellable_executor(embed_mgr.save_to_file)
+            await self._run_cancellable_executor(kg_mgr.save_to_file)
+            
+            return {"status": "success", "count": len(new_raw_paragraphs), "message": f"成功导入 {len(new_raw_paragraphs)} 条知识"}
+            
+        except asyncio.CancelledError:
+            logger.warning("[Plugin API] 导入操作被用户中断")
+            return {"status": "cancelled", "message": "导入操作已被用户中断"}
+        except Exception as e:
+            logger.error(f"[Plugin API] 导入知识失败: {e}", exc_info=True)
+            return {"status": "error", "message": str(e)}
+
+    async def search(self, query: str, top_k: int = 3) -> List[str]:
+        """
+        检索知识库。
+        
+        Args:
+            query: 查询问题。
+            top_k: 返回最相关的条目数。
+            
+        Returns:
+            List[str]: 相关文段列表。
+        """
+        try:
+            _, _, qa_mgr = await self._get_managers()
+            # 直接调用 QAManager 的检索接口
+            knowledge = qa_mgr.get_knowledge(query, top_k=top_k)
+            # 返回通常是拼接好的字符串，这里我们可以尝试按其内部规则切分回列表，或者直接返回
+            return [knowledge] if knowledge else []
+        except Exception as e:
+            logger.error(f"[Plugin API] 检索知识失败: {e}")
+            return []
+
+    async def delete(self, keyword: str, exact_match: bool = False) -> dict:
+        """
+        根据关键词或完整文段删除知识库内容。
+        
+        Args:
+            keyword: 匹配关键词或完整文段。
+            exact_match: 是否使用完整文段匹配（True=完全匹配，False=关键词模糊匹配）。
+            
+        Returns:
+            dict: {"status": "success/info", "deleted_count": 删除条数, "message": "描述"}
+        """
+        try:
+            embed_mgr, kg_mgr, _ = await self._get_managers()
+            
+            # 1. 查找匹配的段落
+            to_delete_keys = []
+            to_delete_hashes = []
+            
+            for key, item in embed_mgr.paragraphs_embedding_store.store.items():
+                if exact_match:
+                    # 完整文段匹配
+                    if item.str.strip() == keyword.strip():
+                        to_delete_keys.append(key)
+                        to_delete_hashes.append(key.replace("paragraph-", "", 1))
+                else:
+                    # 关键词模糊匹配
+                    if keyword in item.str:
+                        to_delete_keys.append(key)
+                        to_delete_hashes.append(key.replace("paragraph-", "", 1))
+            
+            if not to_delete_keys:
+                match_type = "完整文段" if exact_match else "关键词"
+                return {"status": "info", "deleted_count": 0, "message": f"未找到匹配的内容（{match_type}匹配）"}
+
+            # 2. 执行删除
+            # 将同步阻塞操作放到线程池中执行，避免阻塞事件循环
+            
+            # a. 从向量库删除
+            deleted_count, _ = await self._run_cancellable_executor(
+                embed_mgr.paragraphs_embedding_store.delete_items,
+                to_delete_keys
+            )
+            embed_mgr.stored_pg_hashes = set(embed_mgr.paragraphs_embedding_store.store.keys())
+            
+            # b. 从知识图谱删除
+            await self._run_cancellable_executor(
+                kg_mgr.delete_paragraphs,
+                to_delete_hashes,
+                True  # remove_orphan_entities
+            )
+
+            # 3. 持久化
+            await self._run_cancellable_executor(embed_mgr.rebuild_faiss_index)
+            await self._run_cancellable_executor(embed_mgr.save_to_file)
+            await self._run_cancellable_executor(kg_mgr.save_to_file)
+            
+            match_type = "完整文段" if exact_match else "关键词"
+            return {"status": "success", "deleted_count": deleted_count, "message": f"已成功删除 {deleted_count} 条相关知识（{match_type}匹配）"}
+
+        except asyncio.CancelledError:
+            logger.warning("[Plugin API] 删除操作被用户中断")
+            return {"status": "cancelled", "message": "删除操作已被用户中断"}
+        except Exception as e:
+            logger.error(f"[Plugin API] 删除知识失败: {e}", exc_info=True)
+            return {"status": "error", "message": str(e)}
+
+    async def clear_all(self) -> dict:
+        """
+        清空整个LPMM知识库（删除所有段落、实体、关系和知识图谱数据）。
+        
+        Returns:
+            dict: {"status": "success/error", "message": "描述", "stats": {...}}
+        """
+        try:
+            embed_mgr, kg_mgr, _ = await self._get_managers()
+            
+            # 记录清空前的统计信息
+            before_stats = {
+                "paragraphs": len(embed_mgr.paragraphs_embedding_store.store),
+                "entities": len(embed_mgr.entities_embedding_store.store),
+                "relations": len(embed_mgr.relation_embedding_store.store),
+                "kg_nodes": len(kg_mgr.graph.get_node_list()),
+                "kg_edges": len(kg_mgr.graph.get_edge_list()),
+            }
+            
+            # 将同步阻塞操作放到线程池中执行，避免阻塞事件循环
+            
+            # 1. 清空所有向量库
+            # 获取所有keys
+            para_keys = list(embed_mgr.paragraphs_embedding_store.store.keys())
+            ent_keys = list(embed_mgr.entities_embedding_store.store.keys())
+            rel_keys = list(embed_mgr.relation_embedding_store.store.keys())
+            
+            # 删除所有段落向量
+            para_deleted, _ = await self._run_cancellable_executor(
+                embed_mgr.paragraphs_embedding_store.delete_items,
+                para_keys
+            )
+            embed_mgr.stored_pg_hashes.clear()
+            
+            # 删除所有实体向量
+            if ent_keys:
+                ent_deleted, _ = await self._run_cancellable_executor(
+                    embed_mgr.entities_embedding_store.delete_items,
+                    ent_keys
+                )
+            else:
+                ent_deleted = 0
+            
+            # 删除所有关系向量
+            if rel_keys:
+                rel_deleted, _ = await self._run_cancellable_executor(
+                    embed_mgr.relation_embedding_store.delete_items,
+                    rel_keys
+                )
+            else:
+                rel_deleted = 0
+            
+            # 2. 清空知识图谱
+            # 获取所有段落hash
+            all_pg_hashes = list(kg_mgr.stored_paragraph_hashes)
+            if all_pg_hashes:
+                # 删除所有段落节点（这会自动清理相关的边和孤立实体）
+                await self._run_cancellable_executor(
+                    kg_mgr.delete_paragraphs,
+                    all_pg_hashes,
+                    True  # remove_orphan_entities
+                )
+            
+            # 完全清空KG：创建新的空图
+            from quick_algo import di_graph
+            kg_mgr.graph = di_graph.DiGraph()
+            kg_mgr.stored_paragraph_hashes.clear()
+            kg_mgr.ent_appear_cnt.clear()
+            
+            # 3. 重建索引并保存
+            await self._run_cancellable_executor(embed_mgr.rebuild_faiss_index)
+            await self._run_cancellable_executor(embed_mgr.save_to_file)
+            await self._run_cancellable_executor(kg_mgr.save_to_file)
+            
+            after_stats = {
+                "paragraphs": len(embed_mgr.paragraphs_embedding_store.store),
+                "entities": len(embed_mgr.entities_embedding_store.store),
+                "relations": len(embed_mgr.relation_embedding_store.store),
+                "kg_nodes": len(kg_mgr.graph.get_node_list()),
+                "kg_edges": len(kg_mgr.graph.get_edge_list()),
+            }
+            
+            return {
+                "status": "success",
+                "message": f"已成功清空LPMM知识库（删除 {para_deleted} 个段落、{ent_deleted} 个实体、{rel_deleted} 个关系）",
+                "stats": {
+                    "before": before_stats,
+                    "after": after_stats,
+                }
+            }
+
+        except asyncio.CancelledError:
+            logger.warning("[Plugin API] 清空操作被用户中断")
+            return {"status": "cancelled", "message": "清空操作已被用户中断"}
+        except Exception as e:
+            logger.error(f"[Plugin API] 清空知识库失败: {e}", exc_info=True)
+            return {"status": "error", "message": str(e)}
+
+# 内部使用的单例
+lpmm_ops = LPMMOperations()
+