RAGbase Shield Copyright 2026 RAGbase This product includes software developed by RAGbase (https://ragbase.ai). ------------------------------------------------------------------------------- Third-party data shipped with RAGbase Shield ------------------------------------------------------------------------------- src/ragbase_shield/engine/lexicon/first_names_fr.txt Derived from: INSEE, "Fichier des prénoms", édition 2025 (births in France 1900-2025), national file, https://www.insee.fr/fr/statistiques/8595130 Licence: Licence Ouverte / Open Licence 2.0 (Etalab), https://www.etalab.gouv.fr/licence-ouverte-open-licence/ Changes made by RAGbase: names folded (lower case, accents removed), totals summed over sexes and years, names with fewer than 150 births or fewer than 3 letters removed, birth counts dropped. Rebuilt with bench/tools/build_first_names.py. Source: INSEE, Fichier des prénoms, édition 2025. src/ragbase_shield/engine/lexicon/en_first_names.txt Derived from: US Census Bureau, "Frequently Occurring First Names and Surnames From the 1990 Census" (dist.female.first, dist.male.first), https://www2.census.gov/topics/genealogy/1990surnames/ Licence: US government work, public domain. Changes made by RAGbase: names folded, a sex column added (F/M/U from the two files), short and rare names removed, a hand-written supplement of modern and international first names and of nickname pairs added, a few entries removed by the evaluation audit. The header of the file documents the derivation. Every other lexicon file under src/ragbase_shield/engine/lexicon/ (stop words, vocabulary, places, acronyms, legal forms, legal terms, whitelists, NER stop lists, synthetic name, street and town lists) was written for RAGbase Shield. ------------------------------------------------------------------------------- Optional NER layer (v0.2): downloaded by the user, never shipped ------------------------------------------------------------------------------- Installed with: pip install "ragbase-shield[ner]" ; downloaded with: ragbase-shield ner download. Nothing below is in the package. fastino/gliner2-privacy-filter-PII-multi (model weights) Apache-2.0 urchade/gliner_multi_pii-v1 (model weights) Apache-2.0 microsoft/mdeberta-v3-base tokenizer files (urchade model) MIT gliner2, gliner, transformers, huggingface_hub, tokenizers, safetensors, sentencepiece Apache-2.0 torch (CPU build) BSD-3-Clause protobuf BSD-3-Clause numpy, sympy, networkx, mpmath, jinja2, markupsafe, fsspec, filelock, regex, tqdm, requests, urllib3 BSD / MIT / Apache-2.0 (tqdm: MPL-2.0 + MIT) No GPL, LGPL or AGPL component. The models run offline (HF_HUB_OFFLINE=1). ------------------------------------------------------------------------------- Design references (ideas only; no code or data copied) ------------------------------------------------------------------------------- RAGbase Shield was written from scratch. These open-source projects informed its design. We thank their authors. PII Shield (MIT), https://github.com/gregmos/PII-Shield Tools read files on disk and give the model only masked text; a session id stored in the output document allows restore weeks later. PrivAiTe (BSD-3-Clause), https://github.com/crp4222/PrivAiTe A "verify" command that plants test data and reports LEAK or CLEAN. pii-mcp (MIT), https://github.com/foro-sh/pii-mcp Fail closed: on error, never fall back to the raw text. ------------------------------------------------------------------------------- Runtime dependencies ------------------------------------------------------------------------------- RAGbase Shield depends on, but does not bundle, third-party packages that keep their own licences (MIT, BSD, Apache-2.0, PSF-2.0, MPL-2.0). See pyproject.toml. v0.3 adds pdfplumber and pdfminer.six (both MIT) for the PDF table-cell reader; pypdf stays the fallback. The Claude Desktop bundle (.mcpb) contains RAGbase Shield code only; uv installs the dependencies on the user's computer. ------------------------------------------------------------------------------- Benchmark sets (evaluation/datasets) ------------------------------------------------------------------------------- The documents and the answer keys of the REF and Max sets are written by RAGbase. Licence: CC BY 4.0, https://creativecommons.org/licenses/by/4.0/ The method and the case types of matters A, B and C of the Max set follow the white paper Confidentiality and legal artificial intelligence (AIDocs Studio, October 2026). RAGbase wrote every document and every answer key. The Word files were built with python-docx (MIT, Copyright (c) 2013 Steve Canny) and keep parts of its default template. The PDF files embed subsets of the Monaco font (Apple).