Document-RAG

A modern, containerized, end-to-end Retrieval-Augmented Generation (RAG) system for document Q&A.

Motivation

Building a robust RAG system involves more than just a script. It requires:

• Reliable Ingestion: Handling file uploads and chunking them intelligently.
• High-Quality Retrieval: Using state-of-the-art embedding models (bge-m3) and vector databases (Qdrant).
• Precision: Re-ranking results (bge-reranker) to ensure the LLM gets the best context, reducing hallucinations.
• Scalability: Decoupling the heavy ML inference from the lightweight application logic.

This project demonstrates a production-ready architecture for such a system.

Project Structure

• backend/: FastAPI application for orchestration. Managed with uv.
• ml-api/: Dedicated microservice for Embeddings and Reranking. Managed with uv.
• frontend/: React/Vite/Tailwind UI.
• models_cache/: Shared volume for storing downloaded ML models.
• qdrant_data/: Persistent storage for the vector database.
• uploads/: Storage for uploaded documents.

Documentation

• Quickstart Guide: Learn how to set up and run the system (Docker & Local).
• Architecture: Deep dive into the system design, data flow, and stack choices.

Key Features

• Modern Stack: Python 3.10+, React 18, FastAPI, Docker.
• Efficient Dependency Management: Uses uv for lightning-fast, reproducible Python environments.
• GPU Acceleration: ml-api is optimized for CUDA but degrades gracefully to CPU.
• Interactive UI: Clean, responsive chat interface.
• Unified ML API Endpoint: Embedding and reranking use the same service (:8001).

Developer Command Surface

Use the root Makefile for consistent local/CI commands:

make bootstrap
make lint
make test
make docker-build
make docker-up
make docker-smoke
make down
make clean

License

MIT