WeKnora is an open-source, self-hostable platform from Tencent that turns a pile of raw company documents into three things at once: a chat-style search that answers with citations, an autonomous agent that can run multi-step tasks, and an automatically generated wiki. It is aimed at developers and IT teams who have a shared drive full of PDFs, Word files and spreadsheets, and who want an in-house assistant that reads them without the files leaving the company network. The project is MIT-licensed and meant to be deployed on your own servers rather than used as a hosted service.
What it does
You feed WeKnora documents — the README lists PDF, Word, Excel, images and XMind among 10-plus supported formats — and it builds one knowledge base that three different surfaces then sit on top of:
- Question answering. You ask in plain language and get an answer with real citations back to the source documents, instead of a model guessing from memory.
- An agent. A ReAct-style autonomous agent can use tools and sandboxes to carry out multi-step tasks rather than only answering a single question.
- A wiki. The platform generates and maintains a wiki from the same material, with knowledge-graph links between entries, so the documents become browsable rather than only searchable.
The point of the design is that all three read the same index. You do not ingest your files once for search and again for the agent.
How it works
Documents are parsed and indexed into a vector database; a question retrieves the relevant passages, and a language model writes the answer over them with references to where each piece came from. WeKnora supplies the pipeline, the storage wiring and the web interface — it does not ship the intelligence.
That is why the connector list matters. The README says it supports 27 LLM providers, including OpenAI, DeepSeek, Qwen, Claude, Gemini and Ollama, and seven-plus vector databases: pgvector, Milvus, Weaviate, Qdrant, Elasticsearch, OpenSearch and Tencent VectorDB. You pick one model backend and one vector store and the rest of the stack stays the same.
Getting started
The code is at github.com/Tencent/WeKnora. There is no installer to download — you deploy it, and the README describes three routes: Docker Compose, Kubernetes with a Helm chart, or a single-binary "Lite" mode for a smaller setup. The newest release named in the README is v0.8.2, dated 24 September 2026.
What it needs, plainly:
- A server you control, plus one of the supported vector databases. The material saved here does not list minimum CPU, RAM, disk or VRAM figures, so treat sizing as something you will have to work out yourself.
- A model. The software is free under MIT, but the answers are not. Either connect a paid cloud API key (OpenAI, Claude, Gemini, DeepSeek and the rest of the 27) and pay per token, or run a local model through Ollama on your own GPU and pay in hardware and electricity instead. There is no mode where the intelligence is free.
When to use it / when not
The main catch is that cost and operations both land on you: an LLM bill or a GPU, plus a vector database, containers and upgrades to keep alive. This is infrastructure, not an app you install on a laptop in five minutes.
Use it instead of ChatGPT's file upload or a Custom GPT when the documents cannot leave your network, when you need citations you can audit, or when you want an agent and a wiki built from the same corpus rather than a per-chat upload.
Don't use it if you have a handful of PDFs and a question — uploading them to a chatbot you already pay for is faster — or if nobody on the team wants to own a self-hosted service long term.
Alternatives
For anyone who already knows this space, WeKnora sits beside Dify and RAGFlow: ready-to-run retrieval apps that come with a UI, as opposed to LangChain or LlamaIndex, which are libraries you assemble an app out of yourself. If you want the shortest possible path and do not care where the files live, ChatGPT's own file handling remains the easy answer. WeKnora's distinguishing move against its open-source peers is the bundling — search, agent and wiki over one index — not a claim to better retrieval.
Take WeKnora seriously if you already run servers and have a document pile that legal or security will not let you upload anywhere. You get a complete, MIT-licensed stack with unusually broad model and database support, and you can start on a single box with Lite mode before committing to Kubernetes. The headline numbers — 31,800-plus GitHub stars and the number-one "Repository of the Day" spot on 1 November 2025, with 76 days on the trending page — are the project's own figures, credited in its README to star-history.com; they show how much attention it has pulled, not how well it answers your questions. Nobody here has measured its retrieval quality, so budget a pilot on a real subset of your documents before you promise anything internally.