Who WeKnora is for#
Enterprise teams consolidating scattered document stores
WeKnora auto-syncs from Feishu, Confluence, GitLab, Notion, Yuque, DingTalk, Seafile, and RSS into a single searchable knowledge base. Retrieval answers cite the source passage so employees can verify and navigate to the original document. Multi-workspace RBAC lets separate teams control access without separate deployments.
Skip if:
Your knowledge lives in one system already accessible via its native search and you do not need AI-generated answers with cross-source retrieval.
Developer teams building internal AI tools on a private corpus
WeKnora's built-in MCP Server exposes per-workspace endpoints that Cursor, Claude, and other AI tools can query directly. Scoped API keys and a principal model let teams publish knowledge bases to internal tooling without exposing the full admin interface.
Skip if:
You need a public-facing AI product rather than an internal knowledge layer. The README recommends private-network deployment only.
Knowledge-ops roles maintaining a living company wiki
The wiki mode organizes documents into a navigable graph with version rollback. The agent can perform multi-step maintenance tasks using installed skills, and auto-sync connectors keep source documents updated as upstream systems change. Chunk editing with revision history lets editors correct retrieval content without re-ingesting entire documents.
Skip if:
Your team writes documentation directly in Notion or Confluence and does not need AI retrieval layered on top of it.
AI engineers evaluating RAG pipelines over real document corpora
Langfuse integration traces agent steps, retrieval calls, and token usage per pipeline run. Multimodal embedding support including WeMM-Embedding and Qwen3-VL-Embedding lets engineers evaluate image-vector retrieval alongside text. The 29-provider model catalog makes it straightforward to compare retrieval quality across models by changing a catalog config.
Skip if:
You need a programmatic RAG API without a UI. WeKnora is a full application; pipeline-only use cases are better served by a library like LlamaIndex.
The problem it solves#
Enterprise teams accumulate knowledge across Confluence wikis, Notion pages, GitLab repositories, Feishu documents, and scattered chat threads. Finding a specific answer means opening five tools, skimming through context that was never organized for retrieval, and often giving up and asking a colleague who may or may not remember. The pain compounds as the organization grows; the knowledge exists somewhere, but no one can locate it reliably or cite where it came from.
Commercial knowledge-search products address part of this. But they process and store documents on vendor infrastructure, which puts sensitive internal material outside the organization's control. Per-seat subscription pricing scales costs with headcount in ways that make broad rollout expensive. A self-hosted option removes both constraints, but building one from scratch requires assembling a vector search backend, a document parser, a retrieval pipeline, and a chat interface, which is months of engineering work most teams cannot justify.
How it solves it#
Hybrid Retrieval with Source Citations
Runs semantic vector search and keyword search together, returning answers with inline citations linking back to the source passage in the original document. Supported formats include PDF, Word, Excel, XMind, and images, parsed in-process using the anydoc library without external services.
Autonomous Agent with Sandboxed Skills
Executes multi-step tasks using skills installed from ClawHub, SkillHub, Git, or ZIP archives. Skills run in session-persistent Docker, E2B, or Cube sandboxes with an interactive terminal and graphical desktop. The BrowserSkill extension lets the agent drive a user's own Chrome or Edge for tasks that require a real browser session.
Built-in MCP Server per Workspace
Each workspace exposes per-endpoint Streamable HTTP MCP Server endpoints, scoped by knowledge base, with their own token, rate limit, and tool groups. External AI tools like Cursor and Claude can query WeKnora knowledge bases directly over this protocol without any additional proxy.
Broad Data Source Auto-Sync
Auto-sync connectors cover Feishu wiki, Feishu Drive, Confluence, GitLab, Tencent IMA, Notion, Yuque, DingTalk Docs, Seafile, and RSS feeds. Documents update automatically as upstream sources change, keeping the knowledge base current without manual re-ingestion.
Multi-Workspace RBAC with Audit Log
Access control uses a four-tier role matrix (Owner, Admin, Contributor, Viewer) with per-resource ownership and a per-workspace audit log. Multiple storage instances per workspace let teams partition data by sensitivity or team without separate deployments.
Langfuse Tracing and Observability
Agent steps, token usage, and retrieval pipelines write traces to Langfuse, available as an optional Docker Compose profile. A task-queue dashboard shows worker-pool state and per-job status, giving operators visibility into what the agent is doing and why retrieval returned a given result.
Strengths and trade-offs#
Strengths
- MIT License with No Per-Seat CostWeKnora is MIT-licensed, so self-hosted deployments carry no licensing fee and no usage-based charge. Unlike Glean and similar enterprise knowledge platforms that bill per seat, a WeKnora deployment scales headcount without scaling the licensing bill. Internal commercial use is unrestricted.
- 34K Stars and Active Release CadenceAt 34,020 GitHub stars with v0.8.3 released on 2026-10-14, WeKnora has a large community and a frequent release cadence. The changelog shows meaningful feature additions each release, including multimodal embedding, image gallery search, and MCP server improvements, rather than patch-only updates.
- 29 Built-in LLM Providers Including Local OllamaThe model catalog ships with 29 built-in vendors including OpenAI, DeepSeek, Qwen, Gemini, Hunyuan, MiniMax, Ollama, and LiteLLM, with generated metadata for context window, max output, reasoning, and vision capabilities. Teams can run Ollama locally for a fully offline LLM without modifying the application.
- Four Deployment Modes for Any ScaleDocker Compose covers standard multi-service deployments, a Helm chart handles Kubernetes production clusters, a Lite single binary with SQLite runs on a laptop with no external dependencies, and a Desktop GUI provides a login-free start for individual use. Most organizations find a match without custom infrastructure work.
Trade-offs
- -Multi-Service Deployment Requires Ongoing MaintenanceThe standard Docker Compose deployment runs multiple services: web UI, backend API, vector database, and optional Neo4j, MinIO, and Langfuse. Operators manage environment configuration, network policy, and upgrades for each. The README notes breaking changes between releases and requires reading upgrade notes before updating, which adds operational overhead compared to a managed SaaS product.
- -638 Open Issues on a Young RepositoryWeKnora launched in July 2025 and already has 638 open GitHub issues. The issue tracker is growing faster than it is being resolved. Teams relying on less-common integrations or edge-case features should check open issues before committing, as some reported behaviors may not yet have fixes.
- -Deep Tencent Ecosystem OrientationWeKnora integrates closely with Tencent products: WeCom, WeChat Dialog, Feishu, Tencent IMA, and Tencent Cloud Lighthouse. Teams outside the Chinese enterprise software stack will find these integrations irrelevant. The tool works with Slack, Confluence, Notion, and GitLab, but the native integrations and some default UX patterns reflect a Tencent-first perspective.
WeKnora vs alternatives#
WeKnora vs Glean
Glean and WeKnora address the same core problem, helping teams find answers inside a large multi-source document corpus using AI-powered search. The choice between them comes down to deployment model, cost structure, and data control.
| Feature | WeKnora | Glean |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes (Docker, Kubernetes, Lite binary) | No |
| Pricing | Free (infrastructure costs only) | Per-seat subscription |
| Data location | Your infrastructure | Glean's cloud |
| LLM flexibility | 29 built-in providers, local Ollama supported | Managed by Glean |
| Data sources | Feishu, Confluence, GitLab, Notion, Yuque, DingTalk, Seafile, RSS, and more | Enterprise connectors |
| Agent mode | Yes (sandboxed skills, browser control, MCP) | Glean AI assistant |
| Setup required | Docker Compose or Kubernetes | None (managed SaaS) |
WeKnora is the better choice when data sovereignty is a requirement, since documents stay on your servers, LLM API calls go to your chosen provider, and no document content leaves your network. It is also the better choice for cost-conscious organizations, since there is no per-seat licensing charge and the MIT license permits unrestricted internal commercial use. Teams with engineering capacity to manage a multi-service Docker deployment will find the setup accessible.
Glean is the better choice when your team cannot or does not want to operate infrastructure. Its managed SaaS model means zero deployment overhead and no ongoing maintenance burden. Teams that prioritize time-to-value over data control, or that lack the infrastructure experience to manage a multi-service deployment, will typically reach a working knowledge search product faster with Glean.
Quick start#
Deploy WeKnora with Docker Compose; Docker, Docker Compose, and Git are required.
```bash
git clone https://github.com/Tencent/WeKnora.git
cd WeKnora
cp .env.example .env
docker compose pull
docker compose up -d
```What it's built on#
- Languages
- GoJavaScriptPythonTypeScript
- Frameworks
- Vue
- Tooling
- Webpack
FAQ#
Is WeKnora free to use?
Yes. WeKnora is MIT-licensed, so self-hosted deployments carry no licensing fee. You pay for the infrastructure you run it on and any LLM API keys you configure. A managed cloud path exists through Tencent Cloud Lighthouse and the WeChat Dialog Open Platform, but those are separate paid services outside the open-source project.
What LLM providers does WeKnora support?
WeKnora ships with 29 built-in vendors including OpenAI, DeepSeek, Qwen (Alibaba Cloud), Gemini, Hunyuan, MiniMax, NVIDIA, Zhipu, LiteLLM, and Ollama. The model catalog is generated with context-window, max-output, reasoning, and vision metadata per model. To run a local Ollama model, set OLLAMA_BASE_URL in the .env file and specify the embedding model name following the configuration docs.
How does WeKnora compare to Glean?
WeKnora is a self-hosted alternative; Glean is a managed cloud service with no self-hosting option. WeKnora is MIT-licensed with no per-seat fee; Glean charges a subscription. WeKnora requires your team to operate the deployment; Glean offloads that to its own infrastructure. Teams that need data sovereignty or cost control at scale typically favor WeKnora; teams that want zero infrastructure overhead typically favor Glean.
Can WeKnora run without an internet connection?
The core application can run air-gapped if you configure a local Ollama instance for the LLM and use local storage instead of cloud object storage. The Lite single-binary mode with SQLite has no external service dependencies. In this configuration you lose cloud LLM API access and third-party data source syncs, but retrieval and chat over locally uploaded documents continue to work.
How difficult is the initial Docker Compose setup?
The standard path takes four commands: clone the repo, copy .env.example to .env, pull images with docker compose pull, and start services with docker compose up -d. The web UI is then available at http://localhost, and the official Quickstart guide walks through onboarding with sample data. Kubernetes deployment uses a Helm chart in the helm/ directory and targets teams with existing cluster operations experience.
Similar open-source tools#
Cloudflare Os
Open source AI workspace with sandboxed apps and Gatekeeper security
hister
Private full-text search for your pages and files
jentic-one
Self-hosted API broker that keeps AI agent credentials secure
AnythingLLM
Chat with your documents using any LLM, private and self-hosted
Khoj
Self-hosted AI assistant that searches your notes and web
Hermes Agent
Self-hosted AI agent with persistent memory, multi-channel chat, and model choice across OpenAI, OpenRouter, and custom endpoints.