Docs & rules

dataelement/bishengGitHub

This is a public, read-only report. Striff reads this repository's docs, turns each sentence that makes a claim about the code into a rule, and checks the rule against the code on the default branch. How this works

maina8991a5Last read Nov 11

Is this yours? Install to manage it

Once installed, Striff checks every pull request.

Rules summary
Documents
Rules
143
4
4
4
4
129
129
4
41
9
31
42
2
10
1
1
1
3
3
3
1
2
4
3
3
1
1
2
2
2
2
744 docs
dataelement/bishengOpen repository

dataelement/bisheng — documented rules

143 of 143 rules, printed November 17, 2026.

The sentence in your docs
配置面拒绝:ApprovalScenarioAdminService._assert_route_type_allowed() 在 create_route 与 update_route 两处校验,必审场景配 route_type=pass 答 18119。两处都要,否则「建不了 pass 但能把 flow 改成 pass」。注册表不认识的 scenario_code(管理员手工新建的场景)不受限——它们没做过这个承诺。
ApprovalScenarioAdminService has a _assert_route_type_allowed method
Holds
网关兜底:ApprovalGate._forbids_pass(),匹配到的 pass 路由若属必审场景,当作没匹配到(落 route_missing 异常 + 通知管理员),不是抛错也不是放行。这一道不是冗余:路由行还能从手工 UPDATE、从改动前的备份恢复进来,而对必审场景来说「静默自动通过并上线」是最坏的结果,落异常队列让管理员去修才是对的。
ApprovalGate has a _forbids_pass method
Holds
approval_gate.py PASS 分支 — 调 execute_approval_outbox.delay(outbox_id)
ApprovalGate depends on execute_approval_outbox
Holdsapproval worker
approval/application/app_publish.py::withdraw_app_publish → ApprovalCenterService.withdraw_app_publish → app_publish_withdrawal_repository.py 在一个事务内锁定目标实例、取消任务、写 withdrawn 动作。目标由发布模块基于冻结 tenant/app/deployment/version/instance 生成;发布模块先重验当前 publish,实际操作者计审计。通用 withdraw_instance 仍要求申请人。
ApprovalCenterService depends on AppPublishWithdrawalRepository
Holdsservices repositories
RedisManager 继承 BaseContextManager,提供 Redis 客户端的线程安全单例。支持同步和异步访问模式。
RedisManager extends BaseContextManager
Holdscache context
独立的 LangChain 扩展包,位于 src/backend/bisheng_langchain/,作为单独的 Python 包安装,被主应用 bisheng 导入使用。
something in src/backend/bisheng depends on something in src/backend/bisheng_langchain
Holdsbisheng bisheng_langchain
模型 PageData[T], 字段 data List[T] , total int, 说明 新版分页(推荐使用)
2 rules from this sentence
PageData has data
Holds
PageData has total
Holds
GE --> SG
2 rules from this sentence
GraphEngine uses StateGraph
Holds
GraphEngine uses MemorySaver
Holds
GraphEngine(workflow/graph/graph_engine.py)是工作流执行的核心类,负责将工作流 JSON 数据编译为 LangGraph 状态机并驱动执行。
GraphEngine is in src/backend/bisheng/workflow/graph
Holds
属性 nodes_map, 类型 Dict[str, BaseNode], 说明 节点 ID 到节点实例的映射
GraphEngine has nodes_map
Holds
属性 nodes_fan_in, 类型 Dict[str, List[str]], 说明 记录每个节点有哪些前驱节点
GraphEngine has nodes_fan_in
Holds
属性 nodes_next_nodes, 类型 Dict[str, Set[str]], 说明 记录每个节点的所有后续节点
GraphEngine has nodes_next_nodes
Holds
属性 node_level, 类型 Dict[str, int], 说明 从 START 到每个节点的最长路径层级
GraphEngine has node_level
Holds
属性 condition_nodes, 类型 List[str], 说明 所有互斥分支节点(CONDITION 和选择性交互的 OUTPUT)
GraphEngine has condition_nodes
Holds
属性 graph, 类型 CompiledStateGraph, 说明 编译后的 LangGraph 图
GraphEngine has graph
Holds
GraphState(workflow/graph/graph_state.py)是节点间数据传递的核心机制,每个 GraphEngine 实例持有一个 GraphState 对象。
2 rules from this sentence
GraphState is in src/backend/bisheng/workflow/graph
checks “GraphState(workflow/graph/graph_state.py)”
Holds
GraphEngine depends on GraphState
checks “每个 GraphEngine 实例持有一个 GraphState 对象”
Holds
save_context(content, msg_sender) -- 保存用户或 AI 的消息
GraphState has a save_context method
Holds
get_history_memory(count) -- 获取最近 N 条对话历史
GraphState has a get_history_memory method
Holds
get_history_list(count) -- 获取消息对象列表
GraphState has a get_history_list method
Holds
EdgeManage(workflow/edges/edges.py)管理工作流中节点之间的连接关系。
EdgeManage is in src/backend/bisheng/workflow/edges
Holds
回调方法 on_node_start, 事件数据类 NodeStartData, 触发时机 节点开始执行
2 rules from this sentence
BaseCallback has on_node_start
checks “回调方法 on_node_start”
Holds
BaseCallback depends on NodeStartData
checks “事件数据类 NodeStartData”
Holds
回调方法 on_node_end, 事件数据类 NodeEndData, 触发时机 节点执行完成(含错误原因和日志数据)
2 rules from this sentence
BaseCallback has on_node_end
checks “回调方法 on_node_end”
Holds
BaseCallback depends on NodeEndData
checks “事件数据类 NodeEndData”
Holds
回调方法 on_user_input, 事件数据类 UserInputData, 触发时机 INPUT 节点请求用户输入
2 rules from this sentence
BaseCallback has on_user_input
checks “回调方法 on_user_input”
Holds
BaseCallback depends on UserInputData
checks “事件数据类 UserInputData”
Holds
回调方法 on_guide_word, 事件数据类 GuideWordData, 触发时机 引导语输出
2 rules from this sentence
BaseCallback has on_guide_word
checks “回调方法 on_guide_word”
Holds
BaseCallback depends on GuideWordData
checks “事件数据类 GuideWordData”
Holds
回调方法 on_guide_question, 事件数据类 GuideQuestionData, 触发时机 推荐问题输出
2 rules from this sentence
BaseCallback has on_guide_question
checks “回调方法 on_guide_question”
Holds
BaseCallback depends on GuideQuestionData
checks “事件数据类 GuideQuestionData”
Holds
回调方法 on_stream_msg, 事件数据类 StreamMsgData, 触发时机 LLM 流式 token 输出
2 rules from this sentence
BaseCallback has on_stream_msg
checks “回调方法 on_stream_msg”
Holds
BaseCallback depends on StreamMsgData
checks “事件数据类 StreamMsgData”
Holds
回调方法 on_stream_over, 事件数据类 StreamMsgOverData, 触发时机 LLM 流式输出结束
2 rules from this sentence
BaseCallback has on_stream_over
checks “回调方法 on_stream_over”
Holds
BaseCallback depends on StreamMsgOverData
checks “事件数据类 StreamMsgOverData”
Holds
回调方法 on_output_msg, 事件数据类 OutputMsgData, 触发时机 OUTPUT 节点完整消息输出
2 rules from this sentence
BaseCallback has on_output_msg
checks “回调方法 on_output_msg”
Holds
BaseCallback depends on OutputMsgData
checks “事件数据类 OutputMsgData”
Holds
回调方法 on_output_choose, 事件数据类 OutputMsgChooseData, 触发时机 OUTPUT 节点选择交互
2 rules from this sentence
BaseCallback has on_output_choose
checks “回调方法 on_output_choose”
Holds
BaseCallback depends on OutputMsgChooseData
checks “事件数据类 OutputMsgChooseData”
Holds
回调方法 on_output_input, 事件数据类 OutputMsgInputData, 触发时机 OUTPUT 节点输入交互
2 rules from this sentence
BaseCallback has on_output_input
checks “回调方法 on_output_input”
Holds
BaseCallback depends on OutputMsgInputData
checks “事件数据类 OutputMsgInputData”
Holds
任务 执行工作流, 函数 execute_workflow, 队列 workflow_celery, 说明 首次启动工作流执行
execute_workflow is in src/backend/bisheng/worker/workflow
Holds
任务 恢复工作流, 函数 continue_workflow, 队列 workflow_celery, 说明 用户输入后恢复暂停的工作流
continue_workflow is in src/backend/bisheng/worker/workflow
Holds
RedisCallback(worker/workflow/redis_callback.py)继承 BaseCallback,是异步执行场景下的核心回调实现。它承担以下职责:
2 rules from this sentence
RedisCallback is in src/backend/bisheng/worker/workflow
checks “worker/workflow/redis_callback.py”
Holds
RedisCallback extends BaseCallback
Holdsworker workflow
事件桥接 -- 将引擎回调事件序列化为 ChatResponse,写入 Redis 列表(workflow:{unique_id}:event)
RedisCallback depends on ChatResponse
Holdsworker api
Worker 进程内维护全局字典 _global_workflow 缓存暂停中的 Workflow 对象。当工作流进入 INPUT 状态时,对象保留在内存中以便恢复时跳过重新构建。执行完成或失败时从缓存中清除。
_global_workflow depends on Workflow
Holdsworker workflow
questions:问题列表(JSON 数组)
QAKnowledge has questions
Holds
source:来源标识(0=未知,1=手动,2=审核,3=API,4=批量导入)
QAKnowledge has source
Holds
status:QA 状态(0=禁用,1=启用,2=处理中,3=插入失败)
QAKnowledge has status
Holds
BaseFilePipeline(knowledge/rag/base_file_pipeline.py)封装了文件类型到加载器/转换器的映射关系。run() 方法创建临时目录,根据文件扩展名查找 FileExtensionMap,动态初始化对应的 Loader 和 Transformer 链,然后交由 NormalPipeline 执行。
BaseFilePipeline depends on NormalPipeline
Holdsbase_file_pipeline.py pipeline
KnowledgeFilePipeline(knowledge/rag/knowledge_file_pipeline.py)继承 BaseFilePipeline,绑定 KnowledgeFile 数据库记录,负责从 MinIO 下载源文件并构造文件元数据(文档 ID、知识库 ID、上传者、更新者、时间戳、用户自定义元数据等)。
KnowledgeFilePipeline extends BaseFilePipeline
Holds
转换器 文本分块, 类 SplitterTransformer, 说明 使用 ElemCharacterTextSplitter 进行文本切分
SplitterTransformer depends on ElemCharacterTextSplitter
Holdsbisheng bisheng_langchain
方法 init_knowledge_milvus_vectorstore, 说明 初始化单个知识库的 Milvus 向量存储(异步)
KnowledgeRag has init_knowledge_milvus_vectorstore
Holds
方法 init_knowledge_es_vectorstore, 说明 初始化单个知识库的 ES 存储(异步)
KnowledgeRag has init_knowledge_es_vectorstore
Holds
方法 get_multi_knowledge_vectorstore_sync, 说明 批量初始化多个知识库的向量存储,用于跨知识库检索
KnowledgeRag has get_multi_knowledge_vectorstore_sync
Holds
事件是 Agent 执行过程中向上层传递信息的核心机制。所有事件继承自 BaseEvent,包含 task_id 和 timestamp 两个基础字段。事件定义位于 src/backend/bisheng_langchain/linsight/event.py。
2 rules from this sentence
BaseEvent has task_id
Holds
BaseEvent has timestamp
Holds
事件类型 任务开始, 类名 TaskStart, 触发时机 Agent 开始处理某个任务, 核心字段 name
TaskStart has name
Holds
事件类型 任务结束, 类名 TaskEnd, 触发时机 Agent 完成某个任务, 核心字段 name , status , answer , data
4 rules from this sentence
TaskEnd has name
Holds
TaskEnd has status
Holds
TaskEnd has answer
Holds
TaskEnd has data
Holds
事件类型 需要用户输入, 类名 NeedUserInput, 触发时机 Agent 判断需要人工介入, 核心字段 call_reason , params , step_type="call_user_input"
3 rules from this sentence
NeedUserInput has call_reason
Holds
NeedUserInput has params
Holds
NeedUserInput has step_type
Holds
事件类型 生成子任务, 类名 GenerateSubTask, 触发时机 循环任务拆分出子步骤, 核心字段 subtask (子任务列表)
GenerateSubTask has subtask
Holds
ScheduleCenterProcess — 调度中心进程,继承自 multiprocessing.Process。每个 Worker 进程内部通过 asyncio.Semaphore 控制并发任务数。主循环:获取信号量 -> 从队列阻塞消费 -> 注册任务归属 -> 创建 LinsightWorkflowTask 异步任务 -> 任务完成后释放信号量。
ScheduleCenterProcess depends on LinsightWorkflowTask
Holdsworker.py domain
资源初始化:通过 _managed_execution() 上下文管理器,创建 LinsightStateMessageManager、校验会话状态(防止重复执行)、启动终止监控协程、初始化文件目录并下载用户上传的文件到本地临时目录。
LinsightWorkflowTask depends on LinsightStateMessageManager
Holdstask_exec.py services
会话版本模型,记录一次完整的任务执行过程。位于 src/backend/bisheng/linsight/domain/models/linsight_session_version.py。
LinsightSessionVersion is in src/backend/bisheng/linsight/domain/models
Holds
执行任务模型,记录单个步骤的执行状态和历史。位于 src/backend/bisheng/linsight/domain/models/linsight_execute_task.py。
LinsightExecuteTask is in src/backend/bisheng/linsight/domain/models
Holds
所有客户端继承自 BaseMcpClient 抽象基类,统一实现 get_transport() 方法返回读写流,由基类的 initialize() 方法建立 ClientSession。
4 rules from this sentence
SseClient has a get_transport method
checks “统一实现 get_transport() 方法”
Holds
StdioClient has a get_transport method
checks “统一实现 get_transport() 方法”
Holds
StreamableClient has a get_transport method
checks “统一实现 get_transport() 方法”
Holds
BaseMcpClient has an initialize method
Holds
list_tools():列出 MCP 服务器提供的所有工具。
BaseMcpClient has a list_tools method
Holds
call_tool(name, arguments):调用指定工具并返回 JSON 结果。
BaseMcpClient has a call_tool method
Holds
ClientManager.parse_mcp_client_type() 解析传输类型和连接参数。
ClientManager has a parse_mcp_client_type method
Holds
ClientManager.sync_connect_mcp() 创建对应的客户端实例。
ClientManager has a sync_connect_mcp method
Holds
通过 McpTool.get_mcp_tool() 将每个 MCP 工具包装为 StructuredTool。
McpTool has a get_mcp_tool method
Holds
模型类 Role, 文件 role.py, 用途 角色定义,内置管理员角色(ID=1)和默认角色(ID=2)
Role is in src/backend/bisheng/database/models
Holds
Flow 是系统中最核心的模型,通过 flow_type 枚举统一承载六种应用类型:
Flow has flow_type
Holds
Flow.data 字段以 JSON 格式存储完整的画布定义(包含 nodes 和 edges),创建时会自动校验 JSON 结构。FlowVersion 为每个应用维护独立的版本链,is_current 标记当前生效版本。
2 rules from this sentence
Flow has data
checks “Flow.data 字段以 JSON 格式存储”
Holds
FlowVersion has is_current
Holds
Flow.status 控制应用上下线状态:OFFLINE(1) 表示离线编辑中,ONLINE(2) 表示已上线可用。
Flow has status
Holds
Assistant 独立存储助手的 LLM 配置(model_name、temperature、max_token、system_prompt),通过 AssistantLink 关联表将助手与工具(tool_id)、技能(flow_id)、知识库(knowledge_id)三类资源建立多对多关系。
3 rules from this sentence
AssistantLink has tool_id
Holds
AssistantLink has flow_id
Holds
AssistantLink has knowledge_id
Holds
MessageSession 以 chat_id 为主键,记录用户与应用的一次完整对话会话。包含会话统计字段(like、dislike、copied)和敏感词审核状态。group_ids 以 JSON 数组存储所属用户组,支持按组过滤会话。
5 rules from this sentence
MessageSession has chat_id
Holds
MessageSession has like
Holds
MessageSession has dislike
Holds
MessageSession has copied
Holds
MessageSession has group_ids
Holds
编码 1, AccessType KNOWLEDGE, 说明 知识库读权限
AccessType has KNOWLEDGE
Holds
编码 3, AccessType KNOWLEDGE_WRITE, 说明 知识库写权限
AccessType has KNOWLEDGE_WRITE
Holds
编码 5, AccessType ASSISTANT_READ, 说明 助手读权限
AccessType has ASSISTANT_READ
Holds
编码 6, AccessType ASSISTANT_WRITE, 说明 助手写权限
AccessType has ASSISTANT_WRITE
Holds
编码 7, AccessType GPTS_TOOL_READ, 说明 工具读权限
AccessType has GPTS_TOOL_READ
Holds
编码 8, AccessType GPTS_TOOL_WRITE, 说明 工具写权限
AccessType has GPTS_TOOL_WRITE
Holds
编码 9, AccessType WORKFLOW, 说明 工作流读权限
AccessType has WORKFLOW
Holds
编码 10, AccessType WORKFLOW_WRITE, 说明 工作流写权限
AccessType has WORKFLOW_WRITE
Holds
编码 11, AccessType DASHBOARD, 说明 看板读权限
AccessType has DASHBOARD
Holds
编码 12, AccessType DASHBOARD_WRITE, 说明 看板写权限
AccessType has DASHBOARD_WRITE
Holds
编码 99, AccessType WEB_MENU, 说明 前端菜单栏权限
AccessType has WEB_MENU
Holds
Dataset -- 微调数据集,object_name 指向 MinIO 中的数据文件
Dataset has object_name
Holds
Tag / TagLink -- 标签系统,business_type 区分知识库标签(knowledge_space)与应用标签(application),TagLink 通过唯一约束(resource_id + resource_type + tag_id)防止重复绑定
3 rules from this sentence
TagLink has resource_id
Holds
TagLink has resource_type
Holds
TagLink has tag_id
Holds
MarkTask -- 标注任务定义,包含创建者、关联应用 ID、标注人员列表(process_users),状态枚举为 DEFAULT(1)/DONE(2)/ING(3)
3 rules from this sentence
MarkTask has app_id
Holds
MarkTask has process_users
Holds
MarkTask has status
Holds
MarkRecord -- 标注记录,关联 task_id 和 session_id,追踪每条会话的标注状态
2 rules from this sentence
MarkRecord has task_id
Holds
MarkRecord has session_id
Holds
AuditLog 记录系统中的关键操作行为。system_id 标识操作所属模块(chat/build/knowledge/system/dashboard 等),event_type 记录具体行为(如 create_chat、delete_knowledge、user_login),object_type 标识操作对象类型(work_flow/assistant/knowledge 等),并记录操作者 IP 地址。主键使用 UUID 格式。
3 rules from this sentence
AuditLog has system_id
Holds
AuditLog has event_type
Holds
AuditLog has object_type
Holds
存储引擎 Redis 7.0, 基础设施位置 core/cache/redis_manager.py, 存储内容 配置缓存(100s TTL)、Celery 消息代理、Linsight 会话状态(1h 过期)、分布式锁, 访问方式 RedisManager 上下文管理器
RedisManager is in src/backend/bisheng/core/cache
Holds
存储引擎 Elasticsearch, 基础设施位置 core/search/elasticsearch/, 存储内容 稀疏/关键词索引(BM25 检索),遥测统计数据, 访问方式 EsConnManager 上下文管理器,双实例(业务 + 统计)
EsConnManager is in src/backend/bisheng/core/search/elasticsearch
Holds
存储引擎 MinIO, 基础设施位置 core/storage/minio/, 存储内容 文件对象:上传文档、数据集文件、报告模板、应用 Logo、知识库原始文件, 访问方式 MinioManager 上下文管理器,S3 兼容 API
MinioManager is in src/backend/bisheng/core/storage/minio
Holds
DatabaseConnectionManager(core/database/connection.py)负责管理数据库引擎的创建和连接池配置:
DatabaseConnectionManager is in src/backend/bisheng/core/database
Holds
BaseContextManager[T](core/context/base.py)是所有基础设施管理器的抽象基类,提供线程安全的延迟加载、缓存和生命周期管理:
BaseContextManager is in src/backend/bisheng/core/context, and BaseContextManager is a contract
Holds
ApplicationContextManager(core/context/manager.py)作为顶层编排器,按依赖顺序注册并初始化所有基础设施上下文管理器:
ApplicationContextManager is in src/backend/bisheng/core/context
Holds
core/database/tenant_filter.py,由 DatabaseManager._register_tenant_filter() 幂等注册。
DatabaseManager has a _register_tenant_filter method
Holds
授权侧是硬校验:TenantPermissionSubjectDirectory.canonical_source(tenant/domain/services/f048_permission_subject.py:53-91)对三种主体全部要求 tenant 匹配资源自身的 tenant,不匹配抛 PermissionInvalidResourceError(:82-83)。
TenantPermissionSubjectDirectory has canonical_source, and TenantPermissionSubjectDirectory depends on PermissionInvalidResourceError
Holdstenant common
CapabilityCatalogService 通过 ToolServices(action=use) 投影安全字段;picker_tools 返回全部授权工具的分组及空默认选择。WorkCapabilitiesService 读取相同权限资源与已启用租户技能;历史 publication 表保留审计兼容,发布接口返回 410。执行仍验证权限、参数、工具依赖、内容哈希并记录审计。凭据留在宿主侧执行器。技能中途停用使下载作废。
CapabilityCatalogService depends on ToolServices
Holdsdsh tool
决策 1 — 默认值置空,两个历史值列入黑名单。 bisheng/core/config/settings.py jwt_secret: str = ""。新模块 bisheng/user/domain/services/jwt_secret.py 持有 LEGACY_JWT_SECRETS,命中即视为未配置。
LEGACY_JWT_SECRETS is in src/backend/bisheng/user/domain/services
Holds
决策 2 — 未配置时生成一次、存系统配置表。 ConfigKeyEnum.JWT_SECRET = "jwt_secret",值为 secrets.token_urlsafe(48)。选配置表而非本地文件:api / worker / linsight worker 是不同进程甚至不同节点,必须共用同一份;配置表本来就是它们的共同真相源。首次启动多进程竞争插入时靠 key 唯一约束,输家捕获 IntegrityError 后回读胜者的值。
ConfigKeyEnum has JWT_SECRET
Holds
决策 3 — 解析顺序与缓存策略。 resolve_jwt_secret():yaml 值(非黑名单)→ 进程内缓存的生成值 → 读库 / 生成。yaml 值每次实时读、不缓存(读的是内存对象,零成本;也让测试替换 settings 后立刻生效),只缓存生成值。AuthJwt.__init__ 改为调用它;模块内的 DAO / 数据库导入延迟到真正落库时才发生,避免把数据库包拉进早期导入链。
AuthJwt depends on resolve_jwt_secret
Holds
[x] T002: HandleStreamConverter 实现 文件: src/backend/bisheng/citation/domain/services/citation_handle_service.py 逻辑: design §3 决策 2 测试: T001 通过 依赖: T001
HandleStreamConverter is in src/backend/bisheng/citation/domain/services
Holds
domain/schemas/audit.py:只读 AuditRecord/AuditPage;created_at 为带 UTC 时区时间,actor/target/model 名称可空,主体 ID 保留。before_values/after_values/requested_values 为标量快照。
2 rules from this sentence
AuditRecord is in src/backend/bisheng/dsh/domain/schemas
Holds
AuditPage is in src/backend/bisheng/dsh/domain/schemas
Holds
src/backend/bisheng/tool/domain/services/executor.py — ToolExecutor:_init_mcp_tool 包装 McpTool;:157 code_interpreter 注入 minio 配置
ToolExecutor depends on McpTool
Holdstool mcp_manage
DEV-02 流式能力已通:BishengLLM._stream/_astream 对 19 类服务商全链路流式(llm/domain/llm/llm.py L336-356);OpenAI SSE chunk 组装与 OpenAIChatCompletionReq/Resp schema 可参考 open_endpoints/api/endpoints/assistant.py 与 api/v1/schemas(含伪流式降级、reasoning_content 透传)
2 rules from this sentence
BishengLLM has _stream
checks “BishengLLM._stream”
Holds
BishengLLM has _astream
Holds
143 rules from 13 docs

More open-source reports

Every report →

Other public repositories whose docs Striff read and checked against their code.