Import
Upload a file, paste text, or use Data Room to process multiple documents.
Guide
How Safe-Doc pseudonymizes or anonymizes documents (visually).
Safe-Doc secures external AI usage with a simple process: pseudonymize (or anonymize), analyze, de-anonymize.
A controllable flow: import → detect → replace → review → export (with optional restore).
Overview (1 minute)
Input
Single document, pasted text, or multi-document processing with Data Room.
Protection
Choose replacement type and protection level N1–N2 (N3 on the roadmap Q3 2026) based on risk.
Outputs
Copy-ready text + mapping (if reversible) + residual risk indicators.
1) Import: File, Paste, or Data Room
Restore
Re-inject values locally using mapping_xxx.json and a tokenized AI answer.
Data Room
Multi-document analysis with consistent pseudonyms across the whole batch.
Import options
- Upload a file: drag & drop or select a document (PDF/DOCX/TXT).
- Paste text: great for an email, an excerpt, a note.
- Data Room (multi-docs): import multiple documents at once for multi-document analysis.
File
Paste
Data Room
Why Data Room helps
With pseudonymization, the same entity remains the same pseudonym across documents.
Doc 1
Doc 2
Carrefour
Carrefour
[ORG_2]
[ORG_2]
Great for diligence, contracts, multi-piece case files.
2) Choose a mode: Anonymize, Pseudonymize, or Fake data
Anonymized
Generic tokens: [PERSON], [LOCATION]…
When you don't need to keep links between occurrences.
Pseudonymized
Numbered, consistent tokens: [PERSON_1], [ORG_2].
Best to preserve narrative consistency across a document set.
Fake data
Readable replacements (invented names/addresses) for a "natural" text.
Useful for review and presentation while masking real values.
3) Increase protection: N1–N2, N3 (roadmap)
Principle
Higher levels reduce re-identification through context (dates, amounts, locations, writing style).
N1–N2
Standard · Advanced
Direct identifiers (PII) and risky context: tokenization, cleanup and generalization (dates, amounts, locations, references).
N3
High security
Stylistic fingerprint and weak signals. Roadmap Q3 2026.
Visual example
"-17.3M – Feb 12, 2026 – Rouen"
↓
N1–N2: contextual reduction · N3 (roadmap)
↓
"mid-teen millions – Q1 2026 – North France"
Goal: keep useful meaning while reducing identifiability.
4) Review: detected entities and residual scan
Detected entities
Review, filter and adjust what must be masked.
Copy-ready
Copy the pseudonymized result and export mapping/report if needed.
What you control
- Detected entities: people, orgs, locations, emails, phones, IBAN, amounts…
- User choice: uncheck what you don't want to mask.
- Comparisons: views by source (e.g., model vs regex) depending on UI.
Why human review matters
Automated detection can produce false positives and false negatives. Indicators (residual scan / leakage score) help assess risk, but don't replace a final review.
5) Export + Restore (optional)
Primary outputs
- Pseudonymized text (or anonymized depending on mode): ready to paste into an AI tool.
- JSON mapping: original → token mapping (reversible mode).
- Report: audit-friendly indicators (depending on level).
Restore an AI answer
Paste the AI answer containing tokens, then import your mapping to re-inject values locally.
Ready to try it on a real document?