MC3 RAG: GPU-Optimized Document QA

This project is a containerized Retrieval-Augmented Generation (RAG) system developed for AMD Mini Challenge 3. It is designed to answer questions using information retrieved from a supplied document collection and provide evidence-based responses with exact source citations. The system processes the available documents, retrieves relevant information, and generates structured answers for evaluation queries. It is designed to follow the challenge requirements for reliable document-grounded question answering, including handling cases where information is missing or documents cannot be read. In such cases, the system should avoid unsupported claims and return empty answers and citations when required. The solution is packaged as a Docker image based on the provided ROCm/PyTorch environment, targeting deployment and evaluation on AMD GPU infrastructure. The implementation focuses on reproducibility, efficient retrieval, structured JSON outputs, and compliance with the challenge's runtime and resource constraints. The goal is to demonstrate a practical, portable RAG pipeline that combines document retrieval and answer generation while maintaining traceable citations and predictable outputs during automated evaluation.