GraphRAG-Augmented Vietnamese Legal Chatbot

A Graph-enhanced Retrieval-Augmented Generation system for grounded, explainable
and multi-hop legal reasoning on Vietnamese law texts

---

✨ Key Features

• GraphRAG Architecture

  • Hybrid retrieval: Vector Search + Knowledge Graph traversal
  • Neo4j used as a true hybrid database (graph + vectors)

• Rule Engine & Domain Classification

  • Smart Routing: Classifies queries as Criminal, Civil, or Both using regex heuristics
  • Entity Extraction: Pre-computes Articles (Điều), Penalties, and Monetary values to aid Graph traversal
  • Context-Aware: Dynamically injects the correct legal code (BLHS vs BLDS) context into the LLM

• Legal-Aware Processing

  • Hierarchical chunking tailored to Vietnamese legal structure (Chương → Điều)
  • Schema-driven entity & relation extraction
  • Strong emphasis on grounding and hallucination avoidance

• Explainability First

  • Graph-enriched context injected into the LLM
  • Optional graph visualization endpoint
  • Designed to support why an answer is given, not just what

• Dockerized & Reproducible

  • One-command setup
  • Local-first, cloud-optional (AuraDB supported)

---

📦 Tech Stack

• Backend: Python, Flask, LangChain
• Database: Neo4j (AuraDB or local)
• LLMs: Gemini 2.5 Flash Lite (Q/A + extraction)
• Reranking: Cohere Rerank v3.5
• Embeddings: Vietnamese law-specific embedding model (DEk21_hcmute_embedding)
• Deployment: Docker & Docker Compose

---

🤔 How to use?

1. Create the .env File

Create a .env file in the project root and add the following environment variables:

NEO4J_URI=NULL # local or on Aura, your choice
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=password # you can change this
NEO4J_DATABASE=neo4j

GOOGLE_API_KEY=your-google-api-key-here # get api key here https://aistudio.google.com/app/api-keys
COHERE_API_KEY=your-cohere-api-key-here # get api key here https://dashboard.cohere.com/api-keys

VECTOR_INDEX_NAME=chunk_embedding
FULLTEXT_INDEX_NAME=entity_text
ALLOWED_ORIGINS=*

Replace the placeholders with your actual keys.

---

2. Download Embedding Model

If you want to use a local (offline) embedding model, download it from the following link:
👉 https://huggingface.co/huyydangg/DEk21_hcmute_embedding

After downloading, set the configuration as follows:

LOCAL_MODEL_PATH = "/app/models/DEk21_hcmute_embedding"
# True  → Use Local (Offline) Embedding
# False → Use Google Embedding
USE_LOCAL_EMBEDDING = True

3. Build and Start Docker

Ensure Docker is installed on your system, then run:

docker-compose up -d --build

This will build and start all required services.

---

4. Usage Guide

After Docker finishes building, wait a short moment for services to initialize.

You can view container logs using:

docker logs <container_name_or_id>

Once everything is ready, open your browser and navigate to:

http://localhost

to start interacting with the chatbot.

---

⚠ Intended Use & Disclaimer

• This is not a production-ready legal system

• This is a research / academic project

• Designed to study:

  • Hallucination control

  • Legal grounding

  • Graph-augmented reasoning

• Human-in-the-loop is mandatory for any real legal use.

• If your chatbot confidently invents laws, this project exists to prove why that’s unacceptable.

---

👥 Contributors

Thanh Dan Bui Project Manager

Nguyen Dan Vu Backend Developer

Tien Dung Pham Frontend Developer

---