Tencent's Knowledge Platform Stopped Being RAG and Became an Agent
Tencent's WeKnora picked up about 1,200 stars in a day on its way to 25,000, and the reason is that v0.8.0 is no longer a retrieval system with a chat box on it. It has a ReAct agent that orchestrates retrieval, tools, sandboxes and web search across multiple steps, and a Wiki mode where agents write a structured, interlinked Markdown wiki out of your documents, with revision history and human editing on top. MIT, repo at https://github.com/Tencent/WeKnora.
Wiki mode is the interesting bet. Everyone else answers questions on demand and throws the reasoning away. WeKnora materializes the knowledge base into pages that persist, link to each other, and can be corrected by a person who knows better. That turns the corpus into an artifact your team actually reads instead of an oracle you have to interrogate one question at a time, and it means a wrong answer gets fixed once rather than re-generated forever.
The rest reads like someone deployed this inside a large company and got yelled at. Skill sandboxes are session-persistent on Docker, E2B or Cube with per-tenant network policy. Long-term memory carries profiles, preferences and facts across sessions. Four-tier RBAC with per-resource ownership and audit logging, workspace-scoped API keys with capability-level control. Import auto-syncs from Feishu, GitLab, Notion, Yuque and RSS, and twenty-plus model providers are supported including OpenAI, DeepSeek, Qwen, Claude and Ollama.
Everything swaps: vector DB, storage, model provider, and it runs local, Docker, Helm or private cloud. That combination of MIT licence, full data sovereignty and Feishu-and-Yuque connectors tells you exactly who this is for, and it is a market where nobody is going to ship their internal documents to a US API. Worth watching whether the Wiki-as-output idea gets copied by the Western RAG vendors, because it is the best answer anyone has given to the question of what a knowledge base should leave behind.
← Back to all articles
Wiki mode is the interesting bet. Everyone else answers questions on demand and throws the reasoning away. WeKnora materializes the knowledge base into pages that persist, link to each other, and can be corrected by a person who knows better. That turns the corpus into an artifact your team actually reads instead of an oracle you have to interrogate one question at a time, and it means a wrong answer gets fixed once rather than re-generated forever.
The rest reads like someone deployed this inside a large company and got yelled at. Skill sandboxes are session-persistent on Docker, E2B or Cube with per-tenant network policy. Long-term memory carries profiles, preferences and facts across sessions. Four-tier RBAC with per-resource ownership and audit logging, workspace-scoped API keys with capability-level control. Import auto-syncs from Feishu, GitLab, Notion, Yuque and RSS, and twenty-plus model providers are supported including OpenAI, DeepSeek, Qwen, Claude and Ollama.
Everything swaps: vector DB, storage, model provider, and it runs local, Docker, Helm or private cloud. That combination of MIT licence, full data sovereignty and Feishu-and-Yuque connectors tells you exactly who this is for, and it is a market where nobody is going to ship their internal documents to a US API. Worth watching whether the Wiki-as-output idea gets copied by the Western RAG vendors, because it is the best answer anyone has given to the question of what a knowledge base should leave behind.
Comments