TABULA: Grounded OCR and RAG on AMD ROCm

TABULA is a pair of GPU-served document-intelligence systems built for AMD ROCm. TABULA OCR reads licence plates and road signs from PNG, JPEG or TIFF images: a vision-language model reads four pre-processed views of each image, and deterministic code applies the transcription rules (dropping state banners, keeping Chinese province characters, joining lines) and votes across the views. TABULA RAG answers questions over a mixed corpus of PDF, DOCX, XLSX, CSV, logs, code and images. It returns the value only and cites the exact set of files the answer needed, with a grounding check that refuses anything the corpus does not support, and it skips encrypted or unreadable files instead of failing. Both run as one container on the mandated ROCm base image with a vLLM-served Qwen2.5-VL model, load once at start-up, always write a valid output file, and are covered by unit tests, subprocess self-checks and CI. Source: github.com/SriRamkunamsetty/tabula-ocr and tabula-rag.