Skip to content

Veil 0.1.0

Choose a tag to compare

@eleonel eleonel released this 10 Nov 12:30
52b9313

First public release of Veil, a privacy-preserving proxy for LLMs. It detects personal data in text, replaces it with tokens ([PERSON_1]) or fake values, sends the anonymized text to the model, and restores the originals in the reply.

You:      "My SSN is 123-45-6789, help with taxes"
LLM sees: "My SSN is [SSN_1], help with taxes"
You see:  "For your SSN 123-45-6789, I recommend..."

Install

pip install "git+https://github.com/SpinettaInc/veil.git@v0.1.0"
python -m spacy download en_core_web_sm

Python 3.10 or newer. Optional extras: faker, rarity, desktop, full.

What's in it

  • Detection: spaCy NER plus validated regex patterns (e-mail, phone, SSN, credit cards, IBAN, IPv4/IPv6, medical record numbers, dates, addresses), with optional Presidio in hybrid mode. Loose patterns need a cue word nearby.
  • Profiles: paranoid, balanced and minimal, defined in YAML. Custom profiles can change weights and add their own regex detectors and token labels.
  • Replacement modes: token, faker and semantic. One person keeps one token across spellings, titles and partial names.
  • Reconstruction tolerates tokens the model rewrote (EMAIL_1, [email 1], <Email_1>).
  • LLM proxy (VeilProxy) for OpenAI, Anthropic and Ollama: keeps the anonymized conversation itself, anonymizes the system prompt and any supplied history, and streams already-reconstructed text.
  • Fail closed: if a requested detector cannot load, Veil raises instead of anonymizing less. Misspelt profiles and modes are errors too.
  • CLI: veil anonymize, detect, score, reconstruct, stats, serve, app.
  • HTTP API (veil serve) with per-session mappings, a JSONL audit log that records counts and timings only, and batch anonymization.
  • Desktop app (Gradio) and a Guardrails AI validator.

Quality

CI runs lint, the test suite, a labeled precision benchmark in both detection modes with a minimum F1 of 0.95, and a randomised round-trip fuzz. On the benchmark corpus F1 is 0.96 with en_core_web_sm (0.97 in hybrid mode) and 1.00 with en_core_web_lg.

Good to know

  • Only token mode round-trips unambiguously; prefer it for proxy use.
  • The API server has no built-in authentication and binds to 127.0.0.1 by default.
  • Detection is English-only for now.
  • This is an alpha (0.x): the Python API may still change.

Licensed under the GNU General Public License v3.0. Full details are in the changelog and the docs.