Limits and compliance

What Shield does, and what it does not do

Read this page before you trust the result. It describes RAGbase Shield 0.4.4.

What Shield does

Shield does local, reversible pseudonymisation. On your computer, it replaces names and identifiers with codes before the text goes to the AI, then it puts the real names back on your computer.

It is not anonymisation: facts, dates and amounts stay readable by the AI.

Shield reduces what the AI provider receives. It does not replace your judgement or a human review.

Limits (please read)

  1. No OCR.

    A scanned PDF (an image) holds no text. Shield says so and does not read it.

  2. Detection is never complete.

    A rare name, a foreign name, a nickname or a typo can slip through. Read the sensitive passages and the final document again.

  3. Restored names follow the spelling stored in the dossier.

    "M. Mercadier" can come back as "M. Thibault Mercadier", and capitals can change. Identifiers come back in the first spelling seen in the dossier. Read the restored document again.

  4. Facts can identify a client, even without the name.

    A well-known case, a unique job title, a precise place, an amount. Shield does not mask facts. Read again, and generalise when needed.

  5. An admin can block Shield.

    In a firm with a Claude Team or Enterprise account, the administrator can block the adding of Shield. Ask the administrator to allow Shield.

  6. The GDPR still applies.

    For you, the holder of the vault, pseudonymised data remain personal data (GDPR art. 4(5) and recital 26). Shield is a protection measure. It does not release you from your duties.

  7. Text already sent cannot be recalled.

    Shield protects only what goes through it. It does not protect text you type, paste or attach directly in the Claude app, or a folder you add to the conversation yourself.

  8. What Shield cannot stop.

    Malware, or another person with access to your open session. The AI provider's own retention of the masked text: read its terms.

  9. Formats.

    Read: PDF with text, DOCX, XLSX, TXT, MD, EML, HTML. Written: DOCX and MD.

Compliance sheet

This sheet links the texts that matter to a law firm with what Shield does. It is translated from the French sheet. Legal watch of .

Source: the project compliance sheet (docs/CONFORMITE.fr.md, in French).
TextWhat the text saysWhat Shield doesWhat stays with you
GDPR, art. 4(5) GDPR (EUR-Lex) (accessed on ) Pseudonymisation means the data can no longer be attributed to a person without additional information, kept separately and protected. The code-to-name table (the vault) stays on your computer, encrypted. The key is in the OS keychain. Pseudonymised data remain personal data (recital 26). The GDPR applies: register, legal basis, information.
EDPB guidelines 01/2025 on pseudonymisation (version for consultation, adopted on 16/01/2025) EDPB, guidelines on pseudonymisation (accessed on ) Pseudonymisation helps with art. 5, 25 and 32. The data remain personal. The secret that allows re-identification must be protected. Codes given in order (PERSONNE_A, PERSONNE_B), never computed from the name. Vault and key kept apart from the text sent. Your risk analysis.
CNB, practical guide on generative AI systems (2024), p. 31 CNB, practical guide on generative AI (2024) (accessed on ) Pseudonymise before the AI: names replaced by initials or random strings, addresses reduced to the city, care with cross-references (amounts, project names). Names and companies replaced by codes. The whole address masked (ADRESSE_1). Cross-references through facts and amounts.
CNB, AI reading grid, v1.2 (1st edition, June 2025) CNB, AI reading grid v1.2 (2025) (accessed on ) A vendor's promise not to reuse data cannot be checked. Only complete prior anonymisation or a local AI removes the risk. Reduces what the provider receives. Open code, public on GitHub. The risk is not removed: the facts leave. To remove it, you need a local AI.
Paris bar, model AI charter (July 2026), art. 4 and 10 Paris bar, model AI charter (July 2026) (accessed on ) Art. 4, case 2: an online tool is allowed after validation or sufficient safeguards, with no confidential or sensitive data, and a human check every time. Art. 10: security safeguards, including encryption and logging. Open code your referent can audit. Everything stays local. Encrypted vault. Local journal: events, codes and counts, never a real name. Shield masks the identity, not the secret: a fact covered by professional secrecy stays confidential after masking. Validating the AI tool and the review stay with you.
Ethics board of the Avocats au Conseil d'État et à la Cour de cassation, recommendation 2026-01 of 2 April 2026 Avocats aux Conseils, recommendation 2026-01 (accessed on ) Outside a "closed" AI, do not transmit elements that are not anonymised: client names, any information that identifies them, procedure elements. Use pseudonymisation and anonymisation where needed. A legal question asked in the abstract does not breach the secret. Masks names, companies, identifiers and case references (RG, pourvoi, ECLI) before sending. A pseudonymised text is not anonymous. Remove the facts that identify. Prefer the abstract question when you can.

Also worth reading: the CNB guide "Ethics and AI", adopted on 12 and 13 March 2026 (restricted access).

What Shield lets you show

  • The open source code (Apache-2.0). Shield sends nothing to RAGbase.

The 3 rules to give the team

  1. Choose the folder in the Shield window. Do not add it to the conversation.
  2. Do not paste raw text into the chat.
  3. Text already sent cannot be recalled.

This sheet is not legal advice. It describes a tool. Respect for professional secrecy and the GDPR stays the responsibility of the lawyer and the firm.

To remove the risk, a local AI

Shield reduces what the AI sees. It does not replace an AI that runs on your own machines. If your firm wants a private AI installed on its own hardware, RAGbase offers one: ragbase.ai.

See also: frequently asked questions and the results.