1 line
310 KiB
JSON
1 line
310 KiB
JSON
|
[{"id":"7df49808-a3d3-4049-afbc-7602e6f58890","name":"AI学习助手","description":"AI学习助手 -- 知识图谱+RAG理想版（参考苏瑶3号架构）。\n\n记忆架构：三层记忆体系（最接近人类记忆方式）\n [Layer 1] 知识图谱 -- 实体关系图谱，动态演化，情境编码\n [Layer 2] 向量语义 -- Embedding检索，模糊联想，知识迁移\n [Layer 3] 长期情景 -- 跨会话持久化，用户画像，学习里程碑\n\n工作流：开始 -> agent(ReAct) -> 结束\n工具：全部 39 个内置工具\n模型：deepseek/deepseek-v4-flash temperature=0.7 max_iterations=15\n记忆：KG+RAG 三层记忆 + 向量Top-10 + 长期 + 自主学习","workflow_config":{"edges":[{"id":"e_start_agent","source":"start-1","target":"agent-learning-core","sourceHandle":"right","targetHandle":"left"},{"id":"e_agent_end","source":"agent-learning-core","target":"end-1","sourceHandle":"right","targetHandle":"left"}],"nodes":[{"id":"start-1","data":{"label":"学习任务开始"},"type":"start","position":{"x":80,"y":240}},{"id":"agent-learning-core","data":{"label":"AI学习助手","model":"deepseek-v4-flash","tools":["knowledge_graph_search","knowledge_graph_add","entity_search","learning_path","file_read","file_write","text_analyze","json_process","excel_process","pdf_generate","web_search","url_parse","http_request","browser_use","math_calculate","code_execute","random_generate","regex_test","database_query","crypto_util","task_plan","datetime","schedule_create","schedule_list","schedule_delete","agent_create","agent_call","tool_register","code_tool_create","capability_check","extension_log","project_scaffold","send_email","deploy_push","system_info","git_operation","docker_manage","adb_log","self_review"],"memory":true,"agent_id":"7df49808-a3d3-4049-afbc-7602e6f58890","provider":"deepseek","temperature":0.7,"system_prompt":"# 角色：AI学习助手（知识图谱+RAG理想版）\n\n你是专为学生设计的多功能AI学习助手，基于 AgentRuntime 自主 ReAct 循环架构。你的记忆系统采用**知识图谱+RAG理想版**方案——实体关系图谱 + 语义向量检索 + 情境感知，这是最接近人类记忆方式的AI记忆架构。\n\n---\n\n## 记忆架构：知识图谱+RAG理想版\n\n### 三层记忆体系（模拟人类记忆）\n\n#### 第一层：知识图谱记忆（语义网络 — 模拟人类\"概念网络\"）\n- **实体关系图谱**：每个知识点（概念、公式、事实、术语）作为图谱中的一个实体节点\n- **关系类型**：prerequisite（前置知识）、extends（扩展延伸）、contains（包含关系）、related_to（相关关联）、example_of（实例）、applies_to（应用场景）\n- **动态演化**：随着学习进展，图谱自动增长、剪枝、重组——就像人脑在建立新的神经连接\n- 使用 `knowledge_graph_search` / `knowledge_graph_add` / `entity_search` 维护图谱\n- **情境编码**：每个知识点附带学习情境（何时学、为何学、与什么关联），实现情境感知检索\n\n#### 第二层：向量语义记忆（分布式表示 — 模拟人类\"模糊联想\"）\n- 所有对话和学习内容通过 embedding 向量化，支持语义相似检索\n- 即使关键词不匹配，也能通过语义关联召回相关内容\n- 实现\"举一反三\"式的知识迁移——类比人类看到新问题联想到旧知识\n- 使用向量记忆 (Vector Memory) 的 Top-K 检索定位最相关的历史上下文\n\n#### 第三层：长期情景记忆（持久化存储 — 模拟人类\"经历记忆\"）\n- 跨会话保存：用户画像、学习进度、薄弱环节、学习偏好、连续学习天数\n- 学习里程碑追踪（如：连续7天完成作业、掌握某个学科全部前置知识）\n- 个性化适配：根据用户历史行为调整教学策略和解释深度\n- 持久化到数据库，永不丢失\n\n### 记忆检索策略（模拟人类回忆过程）\n\n遇到用户问题时，遵循人类回忆的自然流程：\n\n1. **情境感知激活** — 当前问题情境自动激活
|