Deliverd Screen for Python

A pre-send check for every prompt. Find and redact personal data, secrets and your own confidential terms before text goes to an AI model.

It runs locally, with no account, no key and no network call. It needs no dependencies and works on Python 3.9 and up. MIT.

pip install deliverd-screen

from deliverd_screen import screen

checked = screen("Reply to mary.byrne@example.ie about IBAN IE29AIBK93115212345678")
checked.verdict  # "redact"
checked.text     # "Reply to [EMAIL_1] about IBAN [IBAN_1]"

answer = model.generate(checked.text)
checked.restore(answer)  # the real values, back in the answer

allow sends the original; redact sends placeholders in place of what the model may not receive; ask means a person has to decide; deny sends nothing. By default a model may receive internal data and no personal data. Pass personal_data=True, classification="confidential" for one approved for more.

Names: plug in a model

from transformers import pipeline
from deliverd_screen.adapters import token_classifier, presidio

ner = pipeline("token-classification", model="dslim/bert-base-NER", aggregation_strategy="simple")
checked = screen(prompt, detectors=[token_classifier(ner)])

# or Presidio, in process:
from presidio_analyzer import AnalyzerEngine
checked = screen(prompt, detectors=[presidio(AnalyzerEngine())])

Any model with the same output works, including Perplexity's PII-Tracer and GLiNER. Check its licence before you ship it.

CLI

cat prompt.txt | deliverd-screen --redact
deliverd-screen --fail --json fixtures/*.json

A net, not a certificate

Structured identifiers are found well. Names are found only as well as the model you plug in, and redacted text about a person is still personal data under GDPR.

When a person has to decide, Deliverd's Data Gate asks them. deliverd.screen() in the deliverd SDK runs these same detectors and sends only counts. It takes the same detectors.