Production RAG Engine on AMD ROCm

A high-performance, production-grade Retrieval-Augmented Generation (RAG) engine architected for the AMD AI Academy Challenge (Mini-Challenge 3), achieving a perfect 200/200 score on real AMD GPU hardware with sub-second retrieval latency and 2.94-second corpus indexing. Key Highlights & Architecture: 1. Multi-Format & Multimodal Ingestion: Safely ingests and structures complex heterogeneous enterprise data including PDF, DOCX, XLSX, CSV, LOG, Python source code, and hardware schematics. Integrates an offline RapidOCR pipeline with 2D spatial coordinate mapping to reliably extract dense pinout assignments and printed revision labels without internet access. 2. Deterministic Retrieval & Multi-Hop Reasoning: Combines BM25 token relevance with deterministic entity extraction and multi-hop causal reasoning to bridge diagnostic error logs directly to RMA parts and resolved bug tickets. 3. Necessity Citation Pruning: Employs strict citation pruning to eliminate retrieval noise and false positives, delivering 100% precision. 4. Robust Security & Lifecycle Isolation: Gracefully identifies and rejects obsolete or superseded documents, while providing disciplined refusal on encrypted documents. 5. Strict Hardware & System Compliance: Built on the mandated ROCm 10.0 Ubuntu 26.04 PyTorch runtime, strictly maintaining between 1.0 and 48.0 GiB of active GPU VRAM.