WeKnora: Tencent's Open-Source RAG and Knowledge Framework

What WeKnora is

WeKnora is an open-source, LLM-powered knowledge framework from Tencent for enterprise document understanding, semantic retrieval and reasoning. The README describes its job plainly: bring a team's documents together so they can be searched, reasoned over and kept up to date. The backend is written in Go, the project is MIT-licensed, and it is designed to run on your own infrastructure, with LLMs, vector databases and storage backends all swappable.

What sets it apart from a plain RAG starter kit is that it offers three working modes over the same knowledge bases. Quick Q&A answers questions with retrieval-augmented generation and cites its sources. Smart reasoning runs a ReAct-style agent that plans multi-step work, searches, reads documents and calls tools and skills, showing each step in the conversation. Wiki mode extracts people, products and concepts from documents into interlinked, cited Markdown pages with a knowledge graph. The target audience is teams that want an internal knowledge assistant without sending documents to a third-party SaaS.

How it works

The README describes the architecture as a modular pipeline running from document parsing through vectorization and retrieval to LLM inference, where every component can be replaced or extended. Documents arrive through uploads (folder uploads keep their directory tree) or auto-sync connectors, get parsed (Office files are parsed in-process by a component called anydoc), chunked, embedded and stored in a vector database of your choice. Retrieval combines keyword and vector hybrid search with reranking, parent-child chunking and optional GraphRAG backed by Neo4j.

On top of the retrieval layer sits the agent. Skills installed from ClawHub, SkillHub, Git or ZIP run in session-persistent sandboxes using Docker, E2B or Cube backends, with an interactive terminal and a graphical desktop shown beside the chat. Agents can also connect to external MCP services (including OAuth-protected ones) enabled tool by tool, and since v0.8.2 they can drive the user's own Chrome or Edge through Tencent's open-source BrowserSkill extension, handing control back for logins and CAPTCHAs.

Operationally, WeKnora ships a web UI, a backend API on port 8080, and optional Langfuse tracing. Multi-workspace RBAC, an audit log, scoped API keys and a task-queue dashboard with worker-pool governance cover the multi-tenant side.

Key features

• RAG Q&A with citations: answers drawn from knowledge bases with the sources it used, plus end-to-end evaluation using recall and BLEU / ROUGE.
• ReAct agent: reasons across knowledge bases, web search, MCP tools, skills and the local browser; a running conversation can be steered, forked from an earlier question or rewound with sandbox checkpoints.
• Automatic wiki: agent-generated, interlinked pages with a knowledge graph, in-browser editing, revision diffs and rollback. Release notes say wiki ingest scales to 40k-document knowledge bases.
• Long-term memory: cross-session memory of profile, preferences and facts, auto-extracted but kept only after user confirmation.
• Editable retrieval chunks: chunks can be edited, diffed and rolled back, which helps when a parser gets a table or heading wrong.
• Data source sync: Feishu wiki and Drive, Lark, Confluence, GitLab, Tencent IMA, Notion, Yuque, DingTalk Docs and RSS.
• Broad format support: PDF, Word, PPT, Excel, CSV, TXT, Markdown, HTML, EPUB, MHTML, JSON, XMind and images.
• 27 built-in model vendors: including OpenAI, Azure OpenAI, Anthropic, DeepSeek, Qwen, Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, OpenRouter, LiteLLM and Ollama.
• Many vector stores: PostgreSQL with pgvector, Elasticsearch, OpenSearch, Milvus, Weaviate, Qdrant, Apache Doris and Tencent VectorDB.
• Built-in MCP Server: publishes knowledge bases to Cursor, Claude and other MCP clients over Streamable HTTP, with per-endpoint tokens, scope and rate limits.
• IM channels: Q&A inside WeCom, Feishu, Lark, QQBot, Slack, Telegram, DingTalk, Mattermost, WeChat and Yunzhijia, plus an embeddable website widget.
• Security controls: four-role workspace RBAC, OIDC, AES-256-GCM credential encryption and SSRF-safe outbound requests with a whitelist-only mode.

Getting started

The standard self-hosted path uses Docker Compose and requires Docker, Docker Compose and Git:

git clone https://github.com/Tencent/WeKnora.git
cd WeKnora
cp .env.example .env    # Edit .env as needed, see comments in the file
docker compose pull     # Pull the latest images
docker compose up -d    # Start core services

Then open http://localhost and follow the onboarding guide. The backend API listens on http://localhost:8080 and Langfuse on http://localhost:3000. Optional components are enabled with Compose profiles (full, neo4j, minio, langfuse):

docker compose --profile neo4j --profile minio pull
docker compose --profile neo4j --profile minio up -d
docker compose down     # Stop services

To use a local Ollama model, the README says to run ollama serve > /dev/null 2>&1 & first. Other deployment options are a Helm chart for Kubernetes, a Lite single binary (SQLite plus an in-memory queue, no external dependencies) and a desktop app that currently must be built from source.

There is also an agent-first CLI, weknora, which emits a stable JSON envelope by default:

weknora profile add prod --host https://kb.example.com --use
weknora auth login
weknora kb list
weknora link --kb my-knowledge-base    # bind the current directory
weknora doc upload notes.md
weknora chat "summarise the design doc"

Use cases

• Internal knowledge assistant: sync Confluence, Notion or GitLab into a knowledge base and let staff ask questions from Slack, WeCom or Feishu with cited answers.
• Self-maintaining documentation wiki: point Wiki mode at a pile of specs and meeting notes and get browsable, interlinked pages that can be corrected and rolled back.
• Knowledge for coding agents: expose a knowledge base through the built-in MCP Server so Cursor or Claude can query internal docs while writing code.
• Customer-facing Q&A: publish an agent on an external website through the embed widget, or on WeChat through the hosted WeChat Dialog Open Platform built on WeKnora.
• Multi-step research tasks: use smart reasoning to search the knowledge base and the web, run skills in a sandbox and collect generated files in the artifacts library.

How it compares

The README does not position WeKnora against named competitors. It sits in the same broad category as other self-hosted RAG and knowledge-base platforms that pair document ingestion with a chat UI and agent features. Its distinguishing traits, as described in the README, are the combination of RAG, agent and auto-wiki over one store, the unusually wide list of pluggable backends, and deep integration with Chinese enterprise tooling (WeCom, Feishu, DingTalk, Yuque, Tencent Cloud COS) alongside Western ones like Slack, Confluence and S3. Teams that only need a thin retrieval library to embed in their own application may find WeKnora heavier than necessary, since it is a full application with its own UI, auth and task queue.

Things to know before adopting

• License: MIT. The LICENSE file notes that some listed third-party components are excluded from the MIT grant, so check that list if you redistribute.
• Exposure: the README explicitly recommends deploying in an internal or private network rather than on the public internet, with firewall rules and access controls, despite built-in login.
• Fast-moving releases: v0.8.2 shipped on 2026-09-24 with breaking changes (DingTalk channels are Stream-only, and sandbox commands run as root). Read upgrade notes before pulling new images; docker compose up -d alone reuses cached images.
• Documentation language: the full docs site (around 360 API endpoints and 150 environment variables) is in Chinese, though the README and UI are available in English.
• Hosted vs self-hosted: besides self-hosting, there is a Tencent Cloud Lighthouse template and the hosted WeChat Dialog Open Platform; the code is the same open-source project.
• Sandbox requirements: running skills needs Docker, E2B or Cube; the old local host-process backend was removed in v0.8.0.

Project activity

As of October 2026 the repository has about 31,700 stars on GitHub. It was created on 2025-07-22, is written primarily in Go with a TypeScript/Vue frontend, and is released under the MIT license. The latest release noted in the README is v0.8.2. Source code is at github.com/Tencent/WeKnora and the project site and docs are at weknora.weixin.qq.com.