TaxGPT India 🇮🇳 (Local RAG System)

A production-ready, fully local Retrieval-Augmented Generation (RAG) pipeline designed specifically for Indian Tax Compliance (Income Tax Act, CGST, IGST, ITR Rules). It is built to run 100% locally without any external API keys, ensuring complete data privacy for sensitive financial queries.

Features

• Legal-Aware Chunking: Intelligently splits massive legal PDFs by Sections, Rules, and Chapters to preserve context.
• Source Citations: The LLM explicitly cites the exact source document and page number for every fact [Source 1 | income_tax_act.pdf (Page 45)].
• Query Transformation: Intercepts user queries and rewrites them into optimized vector search queries.
• Microservice Architecture: Decoupled FastAPI backend and Streamlit frontend.
• Local AI Stack: Powered by gemma 4 E4B UD Q4_K_XL.gguf and bge-base-en-v1.5 via FAISS.
• Automated Evaluation: Includes an evaluation pipeline to measure Semantic Similarity, ROUGE-L, and Grounded Precision against a golden dataset of tax queries.

Architecture

1. Ingestion (ingest_data.py): Reads government tax PDFs, performs legal-aware chunking, embeds using SentenceTransformers, and stores in a FAISS index with metadata.
2. Backend (api.py): A FastAPI server with lifespan management and health checks that serves the /chat endpoint.
3. Frontend (app.py): A professional Streamlit web interface tailored for tax queries.

Getting Started

Prerequisites

• Python 3.10+
• Place your local GGUF model in the MODEL/ directory (e.g., gemma-4-E4B-it-UD-Q4_K_XL.gguf).
• Download the official Indian tax laws (Income Tax Act 1961, CGST Act, ITR Instructions) as PDFs and place them in the data/TAX/ directory.

Installation

# Clone the repository
git clone https://github.com/DevilsBreath/Tax_RAG.git
cd Tax_RAG

# Install dependencies
pip install -r requirements.txt

1. Ingest Tax Data

Run the ingestion script to process the PDFs and build the FAISS index:

python ingest_data.py

2. Start the Application

Run the microservices in two separate terminals.

Terminal 1 (Backend):

uvicorn api:app --host 0.0.0.0 --port 8000

Terminal 2 (Frontend):

streamlit run app.py

Docker Deployment

This project is fully containerized.

docker build -t tax_RAG .
docker run -p 8000:8000 -p 8501:8501 tax_RAG

Evaluation

To run the automated evaluation pipeline against the 25-question tax golden dataset:

python scripts/evaluate.py

This will output an evaluation_results_2.md report detailing the performance of the system using ROUGE-L and Semantic Similarity metrics.