Traffic Law Assistant (Vietnamese Legal RAG)

An advanced legal document assistant powered by LightRAG, localized for Vietnamese law and featuring high-fidelity Knowledge Graph visualization. This project uses FastAPI for the backend, React for the frontend, and PostgreSQL (Apache AGE + pgvector) for graph and vector storage.

🚀 Key Features

• Vietnamese Legal Localization: Specialized entity extraction for laws (Điều khoản, Văn bản pháp luật, Cơ quan ban hành).

• Vision-Based PDF Parsing: Uses Qwen 3 VL (via OpenRouter) to extract raw legal text from PDFs with absolute fidelity, even for scans.

• Interactive Knowledge Graph: Explore legal relationships via the integrated LightRAG Graph UI on port 8001.

  

• Comparison Mode: Side-by-side RAG evaluation with parallel streaming.

  

• Hybrid RAG Retrieval: Combined vector and graph search for precise legal grounding.

• Modern Chat Interface: Beautiful React UI with Markdown support and source citations.

• Document Inventory: Manage and track the status of all indexed legal documents.

🛠 Tech Stack

• Backend: Python 3.11, FastAPI, lightrag-hku
• Frontend: Vite, React, TypeScript, Tailwind CSS, Shadcn UI
• Database: PostgreSQL with pgvector (Vector) and Apache AGE (Graph)
• LLM/Embeddings: DeepSeek V3, Qwen 3 VL, OpenAI Embeddings (via OpenRouter)
• Deployment: Docker Compose

📦 Getting Started

Prerequisites

• Docker and Docker Compose
• OpenRouter API Key

Environment Setup

Create a .env file in the root directory (refer to .env.example):

POSTGRES_USER=postgres
POSTGRES_PASSWORD=postgres
POSTGRES_DATABASE=law_assistant
OPENROUTER_API_KEY=your_key_here
LLM_MODEL=deepseek/deepseek-v3.2
EMBEDDING_MODEL=openai/text-embedding-3-small

Running the Application

1. Start the Infrastructure:

   docker compose up -d

2. Start the Frontend (Locally):

   cd frontend
   npm install
   npm run dev

The application will be available at:

• Main UI: http://localhost:5173
• Backend API: http://localhost:8000
• Graph Visualization: http://localhost:8001/webui

🧠 Architecture

The system consists of three main services:

• db: Custom Postgres image with vector and graph extensions.
• backend: Handles chat, PDF parsing, and document indexing.
• rag-ui: Provides the Knowledge Graph visualization interface.

🇻🇳 Localization Details

The RAG engine is optimized for Vietnamese:

• SUMMARY_LANGUAGE: Set to Vietnamese.
• ENTITY_TYPES: Custom legal taxonomy including Hành vi vi phạm, Hình thức xử phạt, Khái niệm pháp lý.

🌍 Recommended Embedding Models

For the best performance with Vietnamese legal text, consider these alternative embedding models:

• Qwen3-Embedding-8B: State-of-the-art multilingual embedding model.
• GreenNode-Embedding-Large-VN-Mixed-V1: Specialized embedding for Vietnamese language tasks.

Note

While models like Qwen3 or GreenNode offer superior performance, OpenAI's text-embedding-3-small (1536D) was chosen as the default for this implementation to stay within the recommended vector dimension limits for efficient pgvector HNSW indexing without excessive memory overhead.

---

🙏 Acknowledgment

Special shoutout to the LightRAG project for providing the powerful Graph RAG framework that powers this assistant.

---

Developed as part of the Traffic Legal Assistant project.