Grounded RAG Container for Mixed Enterprise Docs

We built a Retrieval-Augmented Generation container for AMD AI League Match 3. It is pointed at a folder of mixed documents (PDF, DOCX, XLSX, CSV, TXT/LOG, Python source, PNG/JPG) and answers each question with a short value-only answer plus the exact files it came from. If the corpus does not contain the answer, it returns an empty answer and no citations instead of guessing. The pipeline parses every file type defensively: unreadable, encrypted, unknown and empty items are skipped without stopping the index. Superseded or withdrawn document revisions are tagged and ignored. Images are read with a vision-language model (Qwen2.5-VL), and text is retrieved with hybrid search: BM25, sentence embeddings and an exact-identifier index, fused by reciprocal rank fusion. Multi-hop questions follow identifiers across files (for example, a log naming a ticket, then a database giving the fix version). Answers must be grounded in the retrieved text, and citations are limited to the files the answer actually needs. A resident daemon holds the models and the index, and app.py is a thin client over a unix socket, so each question runs in about a second after warm-up. The container is built on the mandated ROCm PyTorch image, with weights fetched at build time and no network use at run time. Development and testing used a 4 GB NVIDIA GPU; the AMD ROCm build is prepared and awaits testing on AMD hardware.