Unlock Seamless Document Search with WeKnora: An Open‑Source LLM Retrieval Framework

Overview

WeKnora is an open‑source framework from Tencent that combines large language models with retrieval‑augmented generation to enable semantic search and question answering over heterogeneous documents such as PDFs, Word files, and images.

Key Features

Agent mode: Implements ReACT‑style agents that can call built‑in tools, knowledge bases, MCP utilities and web search, iterating to produce comprehensive reports.

Precise understanding: Extracts structured content from PDFs, Word, images and builds a unified semantic view.

Intelligent reasoning: Uses LLMs to interpret document context and user intent, supporting accurate QA and multi‑turn dialogue.

Multi‑type knowledge bases: Supports FAQ and document repositories, with folder, URL import, tagging and online entry.

Flexible extension: Decoupled pipeline (parsing, embedding, retrieval, generation) allows easy integration and customization.

Hybrid retrieval: Combines keyword, vector and knowledge‑graph strategies for cross‑knowledge‑base search.

Web search: Built‑in DuckDuckGo engine for extensible internet search.

MCP tool integration: Extends agent capabilities with MCP, includes uvx and npx launch tools and multiple transport methods.

Conversation strategy: Configurable agent model, normal model, retrieval thresholds and prompts to control multi‑turn behavior.

Simple UI: Intuitive web interface and standard API require no deep technical expertise.

Secure and controllable: Supports on‑premise or private‑cloud deployment; data remains fully under user control.

Technical Architecture

Document processing layer: Parses and preprocesses multi‑format documents (PDF, Word, images).

Knowledge modeling layer: Generates deep knowledge representations via vectorization, chunking and knowledge‑graph techniques.

Retrieval engine layer: Merges keyword, vector and graph retrieval for efficient, precise recall.

Reasoning generation layer: Leverages LLMs for deep understanding and answer generation, with optional agent reasoning.

Interaction display layer: Provides web UI and standard API endpoints.

The modular design allows swapping models (e.g., Ollama with Qwen, DeepSeek) and vector databases, and supports private deployment with firewall protection. Since v0.1.3 a login authentication feature is included.

Quick Start

Prerequisites

Docker

Docker Compose

Git

Installation Steps

1. Clone the repository:

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

2. Copy example environment file and edit variables:

# Copy example env
cp .env.example .env

# Edit .env with required settings (see comments in .env.example)

3. Start the services (includes Ollama):

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

4. Stop the services:

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

Service URLs after launch

Web UI: http://localhost Backend API: http://localhost:8080 Jaeger tracing:

http://localhost:16686

Demo Screenshots

Open‑Source Repository

https://github.com/Tencent/WeKnora