WeKnora: Tencent WeChat Team's Open-Source Enterprise Knowledge Base Unifying RAG, Agents & Auto-Wiki

Overview

WeKnora (维娜拉) is an open-source enterprise knowledge management framework released by Tencent's WeChat team under the MIT license. Built on large language models, it provides a unified knowledge foundation covering the full lifecycle: document ingestion, parsing, knowledge structuring, retrieval-augmented generation (RAG), agent task execution, and continuous knowledge maintenance. Unlike typical RAG tools that only return retrieved fragments for single questions, WeKnora transforms scattered enterprise documents into sustainable, queryable knowledge assets that multiple heterogeneous agents can share.

Three Core Capabilities

1. RAG: Hybrid Retrieval with Editable Chunks

WeKnora combines vector search with BM25 keyword search, fusing results via Reciprocal Rank Fusion (RRF) and re-ranking for higher recall precision. It supports virtually all enterprise file formats (PDF, Word, Excel, PPT, XMind, images, Markdown, CSV, JSON). Crucially, it preserves the original folder hierarchy, allowing users to browse, move, and bulk-tag documents like a network drive. Unlike black-box chunking in most RAG systems, WeKnora lets administrators directly edit individual chunks; if a table or heading is incorrectly split, the chunk can be fixed in-place and the index is automatically rebuilt with version history, diff comparison, and rollback support.

2. ReAct Agent: From Lookup to Task Completion

Traditional RAG answers point queries. WeKnora's built-in ReAct agent autonomously plans steps: deciding when to query the knowledge base, call MCP tools, search the web, or execute scripts in an isolated sandbox. For example, instead of merely retrieving the travel reimbursement policy, the agent can analyze the last three years of policy changes, extract the five most significant updates, and generate a company-wide announcement. Version 0.8.0 introduces a session-level Skill sandbox with three backend options (Docker, E2B, Cube). Administrators can install skills from ClawHub, SkillHub, GitHub, GitLab, or ZIP packages. The agent runs shell commands, reads/writes files, and processes attachments inside the sandbox, returning generated files directly in the chat without contaminating the host environment. Cross-session long-term memory distinguishes user preferences, business facts, and task records; memories persist only after user confirmation and are automatically recalled in later conversations.

3. Wiki Mode: Automatic Knowledge System Construction

The biggest pain point in enterprise knowledge bases is not document scarcity but structural chaos. WeKnora's Wiki mode lets an agent read raw materials, extract entities and relationships, and automatically generate interlinked Markdown Wiki pages alongside a knowledge graph. Pages support manual editing, version control, and rollback. Unlike Obsidian, where graphs link human-authored notes, WeKnora's graph is AI-generated from massive raw business documents, and the resulting Wiki simultaneously feeds both RAG retrieval and agent invocation. This shifts the value proposition from "searching document fragments" to "automatically organizing and continuously maintaining an enterprise knowledge system."

Data Source Synchronization & Multi-Channel Distribution

WeKnora syncs incrementally from Feishu Wiki/Drive, GitLab, Notion, Yuque, Tencent IMA, DingTalk Docs, RSS, and more. Knowledge capabilities can be exposed via web UI, web widgets, CLI, SDK, and MCP Server. IM channels cover WeChat, Enterprise WeChat, Feishu, DingTalk, QQBot, Slack, and Telegram. A Chrome extension saves web content with one click, and the WeChat ClawBot stores shared materials directly into the knowledge base. Most distinctively, WeKnora is the core technical framework for the WeChat Conversation Open Platform, enabling deployed knowledge bases to plug directly into Official Accounts and Mini Programs. This solves the high adoption barrier of standalone AI assistants by leveraging WeChat's mature identity system—users need no extra app or registration.

Model & Storage Decoupling

WeKnora does not lock users into Tencent's Hunyuan model. The model layer is fully pluggable: OpenAI, DeepSeek, Qwen, Zhipu, Hunyuan, Gemini, MiniMax, and local Ollama models are all supported. Chat, embedding, rerank, vision, and speech models can be configured independently. Storage and retrieval components are equally swappable. The default stack uses PostgreSQL with ParadeDB, handling business data, BM25, and pgvector in a single instance for easy startup. Alternatives include Elasticsearch, OpenSearch, Milvus, Qdrant, Weaviate, Doris, and Tencent Cloud Vector Database. Object storage supports local, MinIO, S3, COS, and OSS. Redis manages async tasks, rate limiting, model concurrency, and streaming messages. Enterprise value: switch from DeepSeek to Qwen next year; deploy sensitive workloads on-premises with local models while serving external traffic via cloud models—all without rebuilding the knowledge foundation. Enterprise-grade features include four-level RBAC (Owner/Admin/Contributor/Viewer), fine-grained API key permissions, audit logs, task dashboard, model concurrency limits, and Langfuse tracing.

Horizontal Comparison

Dify : Focuses on AI application orchestration via visual canvas.

RAGFlow : Excels at complex document parsing and high-precision RAG retrieval.

MaxKB : Lightweight, fast delivery of enterprise Q&A bots.

WeKnora : Centers on full-lifecycle knowledge management. A single knowledge asset simultaneously serves RAG Q&A, agent execution, and multi-channel distribution, acting as a unified knowledge layer for multiple heterogeneous agents. It provides an MCP Server so other agents (e.g., DeepSeek Harness, Claude Code) can consume the same private knowledge base.

Local Deployment Tutorial (v0.8.0)

Prerequisites: Docker, Docker Compose, Git.

git clone --branch v0.8.0 --depth 1 https://github.com/Tencent/WeKnora.git
cd WeKnora
cp .env.example .env
# Edit .env to configure models, storage, secrets
docker compose pull
docker compose up -d

Access http://localhost, register an account, and enter the workspace. Recommended onboarding sequence:

Register account, create workspace.

In System Settings, configure chat model, embedding, and rerank models (Ollama local or OpenAI-compatible APIs).

Create a knowledge base, upload enterprise documents.

Wait for parsing, then review and edit chunks, verify citations.

Enable Agent, Wiki, Knowledge Graph, Skill Sandbox as needed.

Optional components are started via profiles to avoid deploying everything at once:

# Knowledge graph requires Neo4j
docker compose --profile neo4j up -d
# MinIO object storage
docker compose --profile minio up -d
# Langfuse tracing
docker compose --profile langfuse up -d

Selection Guide

Choose WeKnora when:

You need long-term maintenance of enterprise knowledge assets, not a one-off Q&A demo.

You want AI to automatically organize documents into structured Wiki and knowledge graphs.

You require private deployment with fine-grained permissions, audit logs, and editable document chunks.

Multiple agents coexist and need a shared enterprise knowledge foundation.

You must integrate with WeChat Official Accounts/Mini Programs and multiple IM channels.

Not recommended when:

Only a quick PDF chat demo is needed → prefer MaxKB.

Core need is visual drag-and-drop AI app orchestration → prefer Dify.

Sole requirement is high-precision retrieval from complex PDFs → prefer RAGFlow.

Conclusion

Two years ago, enterprise AI strategy centered on picking the smartest foundation model. As model capabilities converge, the bottleneck has shifted to private knowledge governance. Enterprise-specific assets—project retrospectives, historical decisions, customer insights—absent from public training data, are the true competitive moat. WeKnora packages RAG, agents, auto-Wiki, knowledge graphs, skill sandboxes, and multi-channel entry points into a single MIT-licensed knowledge foundation. Future enterprises will run many agents built on diverse models and frameworks; WeKnora serves as the central knowledge hub that lets every agent read and understand the organization's proprietary knowledge. The next wave of enterprise AI competition will not be about whose model is smarter, but who can actually deliver scattered enterprise knowledge into the hands of agents.

Project repository: https://github.com/Tencent/WeKnora