Saudi Sign Intelligence – Grounded RAG on AMD

Saudi Sign Intelligence is a citation-exact Retrieval-Augmented Generation container built for AMD Instinct GPUs with ROCm. It answers questions about a folder of mixed enterprise documents with the value only, and cites exactly the files the answer depends on - nothing more, nothing less - or returns an empty answer when the corpus does not hold it. How it works: (1) One resident Qwen3-VL-8B (bf16) server is started by the container CMD and loads once onto the GPU; app.py is a standard-library thin client over a unix socket, so every question process starts in milliseconds and never reloads the model. (2) Defensive parsers for PDF (including scanned pages read by the vision model), DOCX tables in reading order, multi-sheet XLSX, CSV, logs, Python constants and defaults, YAML/TXT and images (OCR by the same VLM). Empty folders, unknown types, unreadable and encrypted files are skipped without stopping the walk. (3) Superseded revisions are detected from what a document says about itself and from revision families, then hidden, so the model never sees the decoy number. (4) Long-context first: when the corpus fits, every current file goes into the prompt; long logs are compressed to their events and rare identifiers instead of being truncated. (5) Citations are verified in code: the value must appear in the cited file, image answers are re-checked by looking at the picture, and chain links (log to ticket to KB article) are kept only when they hand on an identifier the question does not contain. Results: 10/10 on the official sample, 20/20 on a harder self-built stress set, 3-6 s per question. Built in Riyadh by a Saudi Sign Language interpreter and AI developer for trustworthy document Q&A for the Deaf community.