TencentWeKnora is a knowledge framework for document understanding, semantic retrieval, and reasoning over a team's documents. It offers three modes over the same knowledge bases: RAG question answering with citations for lookups, a ReAct agent for multi-step tasks, and an agent-generated wiki for organizing knowledge.
The agent reasons over knowledge bases, web search, MCP tools, skills, and the user's own browser. Skills install from ClawHub, SkillHub, Git, or ZIP archives and run in session-persistent Docker, E2B, or Cube sandboxes, with an interactive terminal and graphical desktop shown beside the chat. Through the BrowserSkill extension the agent operates the user's Chrome or Edge, with a live preview and hand-off for logins and CAPTCHAs. A running conversation can be steered with added requirements, forked from an earlier question, or rewound to a sandbox checkpoint, and cross-session long-term memory keeps profile, preference, and fact entries the user has confirmed.
Documents arrive by upload, by folder (keeping the directory tree), or by auto-sync from sources such as Feishu, Confluence, GitLab, Notion, Yuque, DingTalk Docs, Seafile, and RSS. Retrieval combines keyword and vector search with reranking, parent-child chunking, and GraphRAG on Neo4j, and retrieval chunks can be edited, diffed, and rolled back. LLMs, vector databases, and storage backends are swappable, and WeKnora deploys locally or on a private cloud so the data stays in the operator's environment.
Knowledge reaches users through IM channels (WeCom, Feishu, Slack, Telegram, and others), an embed widget for websites, a built-in MCP server that publishes knowledge bases to MCP clients, an agent-first command-line tool, a Go SDK, a Chrome extension for saving page content, and a WeChat Mini Program. Workspaces carry role-based access control with four roles and an audit log, and Langfuse traces agent steps, token usage, and document-processing runs.
Features
- RAG Q&A: answers from knowledge bases with citations, using hybrid keyword and vector search, reranking, and multimodal parsing
- ReAct agent: multi-step reasoning over knowledge bases, web search, MCP tools, skills, and the local browser
- Wiki: agent-generated, interlinked wiki pages with a knowledge graph, in-browser editing, revision diffs, and rollback
- Skills and sandboxes: skills from ClawHub, SkillHub, Git, or ZIP run in session-persistent Docker, E2B, or Cube sandboxes with per-tenant network policy
- Conversation control: append requirements to a running turn, fork from an earlier question, or rewind to a sandbox checkpoint
- Long-term memory: profile, preference, fact, task, and interest entries extracted across sessions with user confirmation
- Data sources: auto-sync from Feishu, Lark, Confluence, GitLab, Tencent IMA, Notion, Yuque, DingTalk Docs, Seafile, and RSS
- Document formats: PDF, Word, PowerPoint, Excel, CSV, Markdown, HTML, EPUB, MHTML, JSON, XMind, and images
- Backends: vector stores including pgvector, Elasticsearch, OpenSearch, Milvus, Weaviate, and Qdrant, and object storage including MinIO, AWS S3, and several Chinese cloud stores
- Channels and clients: IM bots, a website embed widget, a built-in MCP server, a CLI, a Go SDK, a Chrome extension, and a WeChat Mini Program
- Access control: workspace RBAC with four roles, per-resource ownership, an audit log, scoped API keys, and OIDC sign-in
- Deployment: Docker Compose, Kubernetes via Helm, a Lite single binary with SQLite, or a desktop app
