RAG, Agents, and Wikis All Working on the Same Knowledge Base
You've probably got a pile of documents somewhere—internal specs, meeting notes, support articles—and you've probably tried at least one tool that promised to make them searchable. Maybe it did RAG well. Maybe it could run an agent. Maybe it had some wiki-like organization. But getting all three to play nicely on the same corpus? That's usually where things fall apart. WeKnora, an open-source framework from Tencent, is built around exactly that idea: one set of knowledge bases serving Q&A, multi-step tasks, and organized wikis.
What It Does
WeKnora is an LLM-powered knowledge framework aimed at enterprise document understanding, semantic retrieval, and reasoning. The core premise is straightforward—bring a team's documents together so they can be searched, reasoned over, and kept current. It's released under the MIT license, currently at v0.9.0.
The framework exposes three distinct capabilities that all operate on the same underlying knowledge bases:
- RAG for looking things up, with hybrid search, multimodal parsing, and citations so answers are checkable.
- Agent for multi-step tasks, combining knowledge with tools—skills, a sandbox, a local browser, MCP tools, and memory.
- Wiki for organizing documents into a structured wiki, complete with a knowledge graph and rollback support.
The agent's toolbox is worth noting: skills can be installed from ClawHub, SkillHub, Git, or ZIP files, and run in session-scoped environments. The project also ships clients beyond the core framework—a WeChat Dialog Open Platform integration, a Chrome extension, a ClawHub skill, and an npm package.
Why It's Cool
The shared-knowledge-base design is the real pitch. Plenty of tools bolt RAG and agents together as separate products. Here, the wiki, the retrieval layer, and the agent all read from the same corpus. That means when you organize your documents into a wiki, you're not maintaining a parallel structure that drifts from what the agent actually sees. One source of truth, three ways to use it.
Citations make the RAG side trustworthy. The README emphasizes "answers you can check"—hybrid search plus multimodal parsing plus citations. If you've ever gotten a confident but wrong answer from a document chatbot, you know why this matters. Being able to trace an answer back to its source is the difference between a demo and something you'd actually deploy internally.
The agent gets a genuine toolbox, not just a prompt. Multi-step reasoning is table stakes now. What's more interesting is the combination of skills, a sandbox, a local browser, MCP tools, and memory. The ability to install skills from ClawHub, SkillHub, Git, or ZIP means you're not locked into a fixed feature set—you can extend what the agent can do without forking the project.
Wiki rollback and a knowledge graph suggest maintenance was a design concern. Knowledge bases rot. Documents get updated, structures go stale. Rollback implies the team thought about what happens when someone reorganizes things badly, which is a small detail that says a lot about intended use.
It's genuinely multi-surface. A Chrome extension, a WeChat integration, an npm package, and a ClawHub skill means you can meet the framework where your team already works rather than forcing everyone into a new UI.
How to Try It
The fastest path is to start with the project's own resources:
- Read the docs. The official documentation lives at weknora.weixin.qq.com/docs and is the best starting point for setup specifics.
- Check out the repo. Head to github.com/tencent/weknora for the source, the CHANGELOG, and the README's quick-start section.
- Watch the walkthrough. The README includes a 2:45 video for v0.9.0 with English narration and subtitles—useful for getting a feel for the three modes before you commit to setup.
- Pick your surface. Depending on your workflow, you might start with the Chrome extension, the npm package (
@wxg-prc-cpg/dsh-weknora), or the core framework itself.
The README is written in English, Simplified Chinese, Japanese, and Korean, so if your team spans languages, the onboarding material is already there.
Final Thoughts
WeKnora is aimed at teams that have accumulated real document collections and want more than a single-purpose chatbot—organizations that need retrieval, task automation, and knowledge organization to coexist without maintaining three separate systems. The MIT license and the breadth of client integrations lower the barrier to experimenting. It's early days at v0.9.0, so expect the rough edges that come with a pre-1.0 project, but the core idea—one knowledge base, three ways to use it—is a sensible architecture that solves a real coordination problem. If that problem sounds familiar, it's worth a look.
Follow @githubprojects for more developer tools and open source projects.