Vyson FiscalRAG: cited answers over messy docs
Vyson FiscalRAG is a retrieval-augmented generation container built for AMD AI League Match 3. Pointed at a folder of mixed enterprise documents (PDF datasheets with superseded revisions, DOCX roadmaps with tables, multi-sheet XLSX part lists, CSV ticket databases, production logs, Python source, photographed labels and diagrams), it answers each question with the exact value and the exact set of source files, or refuses when the corpus does not contain the answer.
Architecture: a resident model server starts with the container, loads Qwen2.5-7B-Instruct and Qwen2-VL-2B-Instruct on the AMD GPU (ROCm) and listens on a unix socket, so each per-question process is a thin client well inside the 30-second budget. The index pass walks the corpus defensively (empty directories, unreadable, encrypted and unknown files are logged and skipped, never fatal), parses every supported type, transcribes images with the vision model and persists chunks to disk. Retrieval is BM25 with identifier boosting (ticket ids, error codes, part numbers) plus a second hop that follows rare identifiers across files, which is what makes two-file answers possible. The LLM returns the value and only the passage numbers it strictly needed; citations come from those passages, never from the whole retrieval. Withdrawn or superseded documents are down-weighted and the prompt forbids answering from memory.
Why it matters: the author is a Brazilian accountant and court-appointed expert who spends hours locating the right legal norm, service code and tax rate across hundreds of documents. The same engine powers FiscalRAG, a public Brazilian tax-law assistant over 1,059 legal and municipal records, and is the core of Vyson Fiscal Cloud, a compliance SaaS with eight agents: legislation watcher, tax sheet builder, calculators, tax planner, NFS-e issuer, obligations calendar and auditor for tax recovery. Everything is open source (MIT).
.png&w=1920&q=75)
.png&w=1920&q=75)
.png&w=1920&q=75)
