Tencent Open-Sources WeKnora: RAG Framework for Complex Document Understanding

Introduction

WeKnora is an open-source document understanding and semantic search framework built on large language models (LLMs), designed for structurally complex and content-rich documents. It enables users to move from manual document searching to natural language questioning. The project has garnered over 36,000 stars on GitHub.

Core Features

Agent Mode: Supports ReACT Agent mode, allowing the agent to call built-in tools for knowledge base retrieval, MCP tools, and web search, iterating and reflecting to produce comprehensive summary reports.

Precise Understanding: Extracts structured content from PDFs, Word documents, images, and other formats, building a unified semantic view.

Intelligent Reasoning: Leverages LLMs to understand document context and user intent, supporting accurate Q&A and multi-turn conversations.

Multiple Knowledge Base Types: Supports FAQ and document knowledge bases, with folder import, URL import, tag management, and online entry.

Flexible Extension: Decouples the full pipeline from parsing, embedding, retrieval, to generation, enabling flexible integration and customization.

Efficient Retrieval: Hybrid retrieval strategies combining keyword, vector, and knowledge graph search, with cross-knowledge-base retrieval support.

Web Search: Extensible web search engine with built-in DuckDuckGo integration.

MCP Tool Integration: Extends agent capabilities via MCP, with built-in uvx and npx launch tools supporting multiple transport methods.

Conversation Strategy: Configurable agent model, normal mode model, retrieval threshold, and prompt for precise control of multi-turn dialogue behavior.

Simple and Easy to Use: Intuitive web UI and standard APIs for zero-barrier quick start.

Secure and Controllable: Supports on-premise and private cloud deployment with full data sovereignty.

Technical Architecture

WeKnora adopts a modern modular architecture that cleanly decouples the document understanding and retrieval pipeline:

Document Processing Layer: Handles parsing and preprocessing of multi-format documents (PDF, Word, images, etc.).

Knowledge Modeling Layer: Builds deep knowledge representations through vectorization, chunking, and knowledge graph technologies.

Retrieval Engine Layer: Innovatively fuses keyword, vector, and knowledge graph retrieval strategies for efficient and precise recall.

Reasoning Generation Layer: Uses LLMs for deep understanding and answer generation, integrated with agent reasoning capabilities.

Interaction Display Layer: Provides an intuitive web interface and standard APIs.

This design offers high flexibility: retrieval strategies, LLMs (supports Ollama with switchable models like Qwen, DeepSeek), and vector databases can be freely combined. The end-to-end design ensures controllability and supports private deployment. Since v0.1.3, login authentication has been added; official recommendation is to deploy inside an intranet with proper firewall configuration.

Quick Start

Environment Requirements

Docker

Docker Compose

Git

Installation Steps

1. Clone the repository:

# Clone main repository
git clone https://github.com/Tencent/WeKnora.git
cd WeKnora

2. Configure environment variables:

# Copy example config file
cp .env.example .env

# Edit .env with your configuration
# All variable descriptions are in .env.example comments

3. Start services (including Ollama):

Check the .env file for required images. ./scripts/start_all.sh Or: make start-all 4. Stop services:

./scripts/start_all.sh --stop
# or
make stop-all

Service Access URLs

After successful startup, access:

Web UI: http://localhost Backend API: http://localhost:8080 Distributed Tracing (Jaeger):

http://localhost:16686

Function Demonstration

Web UI Interface

Knowledge Base Management

Conversation Settings

Agent Mode Tool Calling Process

Open Source Address

https://github.com/Tencent/WeKnora