This project implements an enterprise-grade Multimodal Retrieval-Augmented Generation (RAG) system engineered for AMD AI League's Mini-Challenge 3. Built on a mandated AMD ROCm container environment, it answers complex queries across heterogeneous documents—including technical PDFs, Word files, spreadsheets, production logs, source code, and OCR-scanned hardware diagrams. The architecture features a defensive document ingestor resilient to corrupt, encrypted, or permission-restricted files, paired with a sub-millisecond entity-preserving indexer. Operating strictly offline with a resident GPU daemon, the system executes multi-hop reasoning to deliver exact values with strictly necessary citations and calibrated refusals under strict evaluation time constraints.
.png&w=256&q=75)
.png&w=640&q=75)