Index company PDFs and Notion pages and chat with them through an internal RAG service
Run a ReAct agent that uses tools and approval steps to answer harder questions over docs
Generate a connected markdown wiki from a large document set with clickable cross-links
Pipe answers back into WeCom, Feishu, Slack, or Telegram so staff query docs from chat
You need a vector database, an LLM provider key or local Ollama, and a file storage backend wired up before WeKnora can index documents.
WeKnora is an open-source knowledge platform from Tencent that turns piles of documents into something you can chat with, search, and reason over. The basic problem it solves is that most companies have scattered files across many tools, and pulling answers out of them by hand is slow. WeKnora ingests those documents and then offers three ways to use them: quick question-and-answer through a method called RAG, a ReAct agent that strings together multiple steps to handle harder requests, and a new Wiki mode in which agents read the source material and produce a connected markdown knowledge base with a clickable graph of links between pages. The README says it can pull documents in from sources like Feishu, Notion, and Yuque, and that it handles more than ten formats including PDF, Word, Excel, and images. Answers can be served back through chat channels such as WeCom, Feishu, Slack, and Telegram, in addition to the web UI. It is built to plug into many large language model providers, including OpenAI, DeepSeek, Qwen, Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, and Ollama, and the v0.5.2 release notes add Anthropic Claude to that list. A design goal called out repeatedly is that every piece is swappable. You can change the language model, the vector database that stores the document embeddings, the file storage backend, or the search engine without rewriting the rest of the system. The project supports running fully on your own hardware or in a private cloud so that the source documents never leave your network. It also integrates with Langfuse, a tool for watching what the agent is doing inside, tracking token usage, and tracing each step of a pipeline. The latest release notes describe other additions: Wiki ingestion that scales to tens of thousands of documents with a task queue, a human approval step before sensitive tool calls, an adaptive three-tier chunking system that picks how to split each document, a global command palette opened with Cmd-K, a Yuque connector, a small WeChat Mini Program client, and an early command-line client under cli/. There is also a list of bug fixes covering crashes, multi-turn history rebuilds, and OIDC login.
← tencent on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.