n3sfan/NLP-LegalQA

Small models only. Vietnamese legal QA toolkit with scraping, Neo4j/RAG retrieval, LLM QA pipeline - finetune, and a Streamlit UI.

stars

130

commits

Jupyter Notebook

primary language

Jun 6, 2026

updated

evaluation

fastapi

legal-qa

neo4j

rag

streamlit

vietnamese-nlp

README

NLP-LegalQA

NLP-LegalQA is a Vietnamese legal question answering toolkit. It combines legal document scraping, parsing, Neo4j graph import, vector retrieval, reranking, LLM-based answer generation, evaluation scripts, and a Streamlit chat UI.

The project is built for experiments around Vietnamese legal QA and retrieval-augmented generation over structured legal documents.

Features

Scrape Vietnamese legal documents from phapluat.gov.vn.
Parse downloaded documents into structured JSON suitable for graph import.
Import legal articles, clauses, and points into Neo4j.
Generate embeddings and run vector, full-text, hybrid, and reranked retrieval.
Run Legal QA chat through a FastAPI backend and Streamlit UI.
Evaluate retrieval and generation pipelines with scripts for baseline RAG, rerankers, fine-tuned LLMs, BERTScore, ROUGE, and merged evaluation outputs.
Provide annotation tooling for QA references.

Repository Layout

Path	Purpose
`src/legal_scraper/`	CLI, scraper, parser, Neo4j importer, retrieval, API, and QA pipeline code.
`src/llm/`	LLM-related helpers.
`src/annotate_qa/`	FastAPI annotation tool for QA reference data.
`scripts/`	Dataset generation, evaluation, fine-tuning, and result-processing scripts.
`streamlit_ui/`	Streamlit frontend for the Legal QA API.
`tests/`	Unit and integration-style tests for parsing, retrieval, embedding, and graph import behavior.
`notebook/`	Exploratory notebooks and supporting resources.
`qa_dataset/`	QA dataset artifacts used by the experiments.

Requirements

Python 3.12+
Neo4j for graph import, retrieval, and QA pipeline runs
uv or pip
Optional local or hosted OpenAI-compatible LLM endpoint
Optional OpenRouter API key for hosted model experiments

Setup

With uv:

uv sync

Or with pip:

python -m venv .venv
source .venv/bin/activate
pip install -e .

For UI dependencies:

pip install -e '.[ui]'

Copy the environment template:

cp .env.example .env

Configure Neo4j and model provider values in .env.

CLI Examples

Search legal documents:

legal-scraper search "giao thông đường bộ" -n 10

Scrape documents:

legal-scraper scrape "giao thông đường bộ" -n 20 -o data

Parse downloaded documents:

legal-scraper parse -i data -o data/parsed

Import parsed documents into Neo4j:

legal-scraper import-neo4j -i data/parsed

Run vector search:

legal-scraper vector-search -q "quy định về tốc độ xe máy" --k 5

Run the full QA pipeline:

legal-scraper query -q "Người điều khiển xe máy cần giấy tờ gì?"

API and UI

Start the FastAPI backend:

uv run uvicorn legal_scraper.api:app --host 0.0.0.0 --port 8000 --reload

Run the Streamlit UI:

cd streamlit_ui
streamlit run app.py

The UI expects the API backend at http://localhost:8000.

Tests

pytest

Some tests and pipeline commands require local data files, Neo4j, model weights, or API credentials. Keep secrets in .env; do not commit local credentials or generated payloads.

Status

This is an active research/course project for Vietnamese legal QA. The codebase is organized for experimentation and evaluation, not yet as a packaged production service.

License

MIT

Contributors

72 commits

24 commits

21 commits

12 commits

Similar repos

abdullah-ax/safety-advisor-copilot

Personal RAG project answering OSHA workplace-safety questions (LangChain, FastAPI, pgvector). Not…

shared topics

zbilgeozkan/rot-rag-project

The RoT RAG Project implements a Python-based RAG pipeline. Documents are ingested and split into…

shared topics

boostcampaitech5/level3_nlp_finalproject-nlp-08

사용자가 채팅웹을 통해 자신이 처한 법률적 상황을 제시하면, 입력에 대한 문맥을 모델이 이해하여 가이드라인을 제시하고, 유사한 상황의 판례를 제공하는 웹 서비스입니다.…

boostcampaitech5/LawBot-Online-Legal-Advice-LLM-Service

Kush614/sf-intelligence-backend

SF Intelligence Platform - FastAPI backend with RAG, FAISS, vLLM

oncologyxai/CancerRAGent

DocAILab/XRAG

XRAG: eXamining the Core - Benchmarking Foundational Component Modules in Advanced…

305

shared topics

Kingsley-Cyber/polymath_v3.3

Local-first GraphRAG workbench for ingesting large document corpora into vector search and Neo4j…

shared topics

Languages

Jupyter Notebook

52.9%

Python

44.3%

HTML

1.9%

n3sfan/NLP-LegalQA

Small models only. Vietnamese legal QA toolkit with scraping, Neo4j/RAG retrieval, LLM QA pipeline - finetune, and a Streamlit UI.

stars

130

commits

Jupyter Notebook

primary language

Jun 6, 2026

updated

evaluation

fastapi

legal-qa

neo4j

rag

streamlit

vietnamese-nlp

README

NLP-LegalQA

The project is built for experiments around Vietnamese legal QA and retrieval-augmented generation over structured legal documents.

Features

Scrape Vietnamese legal documents from phapluat.gov.vn.
Parse downloaded documents into structured JSON suitable for graph import.
Import legal articles, clauses, and points into Neo4j.
Generate embeddings and run vector, full-text, hybrid, and reranked retrieval.
Run Legal QA chat through a FastAPI backend and Streamlit UI.
Evaluate retrieval and generation pipelines with scripts for baseline RAG, rerankers, fine-tuned LLMs, BERTScore, ROUGE, and merged evaluation outputs.
Provide annotation tooling for QA references.

Repository Layout

Path	Purpose
`src/legal_scraper/`	CLI, scraper, parser, Neo4j importer, retrieval, API, and QA pipeline code.
`src/llm/`	LLM-related helpers.
`src/annotate_qa/`	FastAPI annotation tool for QA reference data.
`scripts/`	Dataset generation, evaluation, fine-tuning, and result-processing scripts.
`streamlit_ui/`	Streamlit frontend for the Legal QA API.
`tests/`	Unit and integration-style tests for parsing, retrieval, embedding, and graph import behavior.
`notebook/`	Exploratory notebooks and supporting resources.
`qa_dataset/`	QA dataset artifacts used by the experiments.

Requirements

Python 3.12+
Neo4j for graph import, retrieval, and QA pipeline runs
uv or pip
Optional local or hosted OpenAI-compatible LLM endpoint
Optional OpenRouter API key for hosted model experiments

Setup

With uv:

uv sync

Or with pip:

python -m venv .venv
source .venv/bin/activate
pip install -e .

For UI dependencies:

pip install -e '.[ui]'

Copy the environment template:

cp .env.example .env

Configure Neo4j and model provider values in .env.

CLI Examples

Search legal documents:

legal-scraper search "giao thông đường bộ" -n 10

Scrape documents:

legal-scraper scrape "giao thông đường bộ" -n 20 -o data

Parse downloaded documents:

legal-scraper parse -i data -o data/parsed

Import parsed documents into Neo4j:

legal-scraper import-neo4j -i data/parsed

Run vector search:

legal-scraper vector-search -q "quy định về tốc độ xe máy" --k 5

Run the full QA pipeline:

legal-scraper query -q "Người điều khiển xe máy cần giấy tờ gì?"

API and UI

Start the FastAPI backend:

uv run uvicorn legal_scraper.api:app --host 0.0.0.0 --port 8000 --reload

Run the Streamlit UI:

cd streamlit_ui
streamlit run app.py

The UI expects the API backend at http://localhost:8000.

Tests

pytest

Some tests and pipeline commands require local data files, Neo4j, model weights, or API credentials. Keep secrets in .env; do not commit local credentials or generated payloads.

Status

This is an active research/course project for Vietnamese legal QA. The codebase is organized for experimentation and evaluation, not yet as a packaged production service.