# WeKnora: Tencent's Open-Source LLM Knowledge Platform > WeKnora is an open-source, Go-based knowledge platform from Tencent that turns documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining wiki. All three modes share the same knowledge bases and can be deployed on-premises with Docker Compose. **Tencent/WeKnora** — Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki. - Repository: https://github.com/Tencent/WeKnora - Website: https://weknora.weixin.qq.com - Stars: 31,206 · Forks: 4,171 - Language: Go - License: NOASSERTION - Published: 2026-09-09 · Updated: 2026-09-09 · Language: en - Canonical page: https://hysenlabs.com/projects/tencent-weknora ## What WeKnora Solves and Who Uses It Enterprise teams accumulate documents in disconnected silos: shared drives, wikis, chat threads, and project management tools. Searching across them is slow and the answers are stale. WeKnora addresses this by ingesting those documents into knowledge bases that three different interfaces can then query: RAG for point lookups, an agent for multi-step reasoning tasks, and a wiki for organized, curated knowledge. All three operate on the same underlying data, so a document indexed once is immediately available to all three modes. The primary audience is engineering and operations teams that need document-grounded answers inside IM tools they already use, such as WeCom, Feishu, Slack, and Telegram. A built-in MCP Server also lets tools such as Cursor and Claude connect to WeKnora knowledge bases directly, making it a back-end knowledge layer for AI coding assistants. ## Three Modes on a Shared Knowledge Base WeKnora exposes three distinct modes, each suited to a different query type. RAG mode handles point lookups. A user asks a question and the system retrieves the most relevant document chunks, returning a grounded answer with citations. This is the standard retrieval-augmented generation loop. Agent mode handles multi-step tasks. The agent can invoke skills installed from ClawHub, SkillHub, Git, or ZIP archives. Skills run inside session-persistent Docker, E2B, or Cube sandboxes with an interactive terminal and graphical desktop beside the chat. Through the BrowserSkill extension the agent can operate the user's own Chrome or Edge browser. External MCP services with OAuth support can be connected and enabled tool by tool. Wiki mode provides curated, organized knowledge. Cross-session long-term memory keeps user profiles, preferences, and confirmed facts persistent across sessions. Folder uploads preserve the original directory tree, and individual retrieval chunks can be edited, compared, and rolled back. The design choice to share knowledge bases across all three modes means a document uploaded for wiki organization is immediately searchable in RAG and accessible to the agent without re-indexing. ## Deploying WeKnora with Docker Compose The standard deployment path uses Docker Compose. It requires Docker, Docker Compose, and Git. ```bash git clone https://github.com/Tencent/WeKnora.git cd WeKnora cp .env.example .env docker compose pull docker compose up -d ``` After the containers start, the web UI is available at http://localhost, the backend API at http://localhost:8080, and the optional Langfuse tracing interface at http://localhost:3000. The .env.example file documents each variable with inline comments. At minimum, teams need to configure the LLM provider, vector database, and storage backend before onboarding documents. WeKnora supports optional service profiles to add capabilities beyond the core stack: ```bash docker compose --profile neo4j --profile minio pull docker compose --profile neo4j --profile minio up -d ``` The neo4j profile adds a knowledge graph backed by Neo4j. The minio profile adds object storage via MinIO. The langfuse profile enables tracing for agent steps, token usage, and pipelines. The full profile activates all features at once. To upgrade, set the target release tag in .env and pull the new images: ```bash docker compose pull docker compose up -d ``` The README notes that running docker compose up -d alone reuses locally cached images and can leave the UI version out of sync with the release images. The upgrade notes for each release document any migration steps required. ## Data Sources, Document Formats, and LLM Vendors WeKnora can auto-sync from Feishu Wiki, Feishu Drive, Confluence, GitLab, Tencent IMA, Notion, Yuque, DingTalk Docs, and RSS feeds, with additional connectors described as in progress. On the document side it handles PDF, Word, images, Excel, and XMind files, among others, parsing Office formats in-process using anydoc. The LLM layer supports 27 built-in vendors including OpenAI, DeepSeek, Qwen (Alibaba Cloud), Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, LiteLLM, and Ollama. LLMs, vector databases, and storage backends are all swappable, which means teams can run WeKnora entirely on-premises with a local Ollama model and a self-hosted vector store without any external API calls. The README gives this example for using a local Ollama model: ```bash ollama serve > /dev/null 2>&1 & ``` Run this before starting WeKnora, then configure the OLLAMA_BASE_URL and embedding model name in the .env file. ## Permissions, Audit Logging, and Langfuse Tracing WeKnora includes multi-workspace RBAC with four roles: ownership and permissions can be set per resource and per workspace. Each workspace gets its own audit log, so teams can trace who queried what and when. On the operational side, a runtime task-queue dashboard provides visibility into worker-pool activity. Langfuse tracing records individual agent steps, token usage, and pipeline execution, making it possible to identify which retrieval call consumed the most tokens or which agent skill caused a task to fail. Scoped API keys with a configurable principal model allow programmatic access with predictable cost controls. This operational detail is meaningful: many open-source knowledge platforms provide no visibility into what the agent actually did between receiving a query and returning an answer. WeKnora's tracing layer exposes that process. ## Limitations and Cases Where WeKnora Is the Wrong Tool WeKnora's multi-service architecture is its main operational overhead. The core stack requires a frontend container, an app container, a database, a Redis instance, a vector store, and optional services for graph, object storage, and tracing. Teams that want a single-binary or minimal-dependency RAG tool will find this heavy. The agent sandbox runs inside Docker, E2B, or Cube. Teams in environments where running Docker-in-Docker or external sandbox services is prohibited will find the agent's skill execution unavailable. The auto-sync integrations lean toward Chinese enterprise tools: Feishu, Yuque, Tencent IMA, and DingTalk Docs are explicitly supported. Teams using only Western tools such as Google Drive, Dropbox, or Sharepoint would need to verify whether those integrations exist before committing. Finally, the license. GitHub reports the license as NOASSERTION, which means the automated detection could not identify the license type. A LICENSE file is present in the repository, but teams should read it directly and consult their legal team before using WeKnora in a commercial deployment. ## How WeKnora Compares to RAGFlow RAGFlow, maintained by infiniflow, is another open-source document RAG platform with a similar goal: ingest enterprise documents and make them queryable. The key difference in approach is scope. RAGFlow focuses on the retrieval pipeline: document parsing, chunking strategies, and retrieval quality. WeKnora bundles a full agent execution environment with skill management and sandbox execution alongside the RAG layer. Teams that need only retrieval-grounded Q&A will find RAGFlow simpler to operate and easier to reason about. Teams that want the agent to take actions, such as filing a ticket, querying an API, or running code inside a sandbox, in response to a retrieved result will find WeKnora's broader toolbox relevant. The wiki mode and IM channel integrations (WeCom, Feishu, Slack, Telegram) are specific to WeKnora and have no direct equivalent in RAGFlow. ## Conclusion Teams that already work inside the WeChat or Tencent ecosystem and need a single platform covering retrieval, agent reasoning, and wiki management will find WeKnora the most integrated option available. Teams running entirely on open infrastructure without Tencent dependencies should first audit how much of the feature set they actually use; a leaner RAG-only stack may suffice. Before deploying, review the LICENSE file directly: the repository's detected license type is NOASSERTION, meaning GitHub could not identify the license automatically. Verify that the terms permit your intended use before running WeKnora in a production environment. ## FAQ ### What LLM providers does WeKnora support? WeKnora includes 27 built-in vendors, including OpenAI, DeepSeek, Qwen (Alibaba Cloud), Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, LiteLLM, and Ollama. Local models via Ollama can be used for fully on-premises deployments. ### Can WeKnora be deployed without an internet connection? The README describes a self-hosted deployment path using Docker Compose where all services run on your own infrastructure. Configuring a local Ollama model and a self-hosted vector store removes the need for external API calls during inference, though the initial image pull from Docker Hub requires internet access. ### What is the difference between WeKnora's RAG mode and its agent mode? RAG mode retrieves document chunks and returns a grounded answer in a single round trip. Agent mode executes multi-step tasks using installed skills that can run code in Docker, E2B, or Cube sandboxes, operate a browser, and call external MCP services, making it suitable for tasks that require more than a lookup. ## Sources - [Issues](https://github.com/Tencent/WeKnora/issues) - [Project website](https://weknora.weixin.qq.com) - [README](https://github.com/Tencent/WeKnora/blob/main/README.md) - [Releases](https://github.com/Tencent/WeKnora/releases) - [Tencent/WeKnora on GitHub](https://github.com/Tencent/WeKnora) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/tencent-weknora