MC3 RAG on AMD MI300X — 200/200

A retrieval-augmented QA container for AMD AI League ACT III Mini Challenge 3, built and verified on an AMD Instinct MI300X (ROCm 10.1). How it works. The image boots a persistent GPU daemon that keeps both models resident, so per-question processes never reload weights. On --index, the corpus (PDF, DOCX, images, spreadsheets) is chunked and embedded with BAAI/bge-small-en-v1.5; images are OCR'd by the vision model. On each query, the retriever pulls the top chunks, a stale-revision penalty demotes superseded documents (e.g. withdrawn datasheet revisions), and Qwen/Qwen2.5-VL-7B-Instruct produces an exact answer plus a verified citation set — every cited file is checked against the corpus before output. Contract. `python3 /app/app.py --index /app/corpus` for indexing, then per-question `python3 /app/app.py --corpus /app/corpus --query-id <id> --query "..."` writes `/app/output/<query-id>_output.json` with fields answer, citations, confidence. Verified results. On the sample corpus: 10/10 questions correct, TOTAL 200/200. Self-check on the MI300X: container starts with no network in 0.1s, index completes in 24.9s, a query answers in 0.4s (30s budget), VRAM stays inside the 1-48 GiB band. The trap question (withdrawn r1 datasheet claiming 105 C vs current r2) is answered correctly: '94 C' cited to specs/tq40_datasheet_r2.pdf. Stack: ROCm 10.1 base image, PyTorch, Python, Docker. Models are baked in at build time, so the container runs fully offline at judging time.