[Gitlab repo | Github mirror]

EasyRedact

AI-assisted PDF redaction tool for German federal ministries (BMZ). Upload a PDF or DOCX, have the AI suggest redactions via an interactive chat, review and adjust them, then export a fully redacted document — everything runs in the browser, documents never leave the machine unless you choose Cloud AI mode.

Features

• Two redaction modes

  • PII — redacts personal data: names, addresses, emails, phone numbers, bank details, dates of birth
  • FOI / IFG — redacts based on the exemption clauses of a chosen Freedom of Information law; jurisdiction rules are loaded at runtime from datenlabor-bmz/redaction-rules

• Cloud AI and Local AI (switchable at any time via the mode selector)

  • Cloud AI — any OpenAI-compatible LLM endpoint (e.g. Azure AI Foundry, GDPR-compliant, no data retention)
  • Local AI — processes documents on your own infrastructure or in the browser; configured via LOCAL_AI (see below)

• AI chat assistant — reads the document, asks targeted clarifying questions, then suggests redactions with exact text matches, confidence ratings, affected persons, and legal justifications

• Manual redactions — draw rectangles or select text directly on the PDF without AI involvement

• Multi-document tabs — open multiple PDFs at once; session (documents, redactions, chat) is persisted in IndexedDB and survives page reload

• Export

  • Preview PDF — yellow highlight boxes for review and sign-off
  • Redacted PDF — text permanently removed via MuPDF, ready for publication

• DOCX → PDF conversion — LibreOffice-backed, Docker deployment only

Architecture

Three-panel layout rendered entirely client-side:

Panel	Content
Left	Redaction list grouped by person/page; accept / ignore controls; FOI rule assignment
Center	PDF viewer (MuPDF WASM in a Comlink worker); zoom; export
Right	Chat or NLP panel; AI mode selector; streaming SSE from /api/chat

The Next.js API routes (/api/chat, /api/docx, /api/nlp) are thin server-side proxies — all document rendering and redaction geometry stay in the browser.

Getting Started

npm install
npm run dev

Open http://localhost:3000. Copy .env.example to .env and fill in at minimum the Cloud LLM credentials.

Environment Variables

See .env.example for the full list. Key variables:

# Cloud LLM (any OpenAI-compatible API — Azure AI Foundry, OpenAI, etc.)
CLOUD_LLM_API_BASE=https://YOUR-RESOURCE.openai.azure.com/openai/v1
CLOUD_LLM_API_KEY=
CLOUD_LLM_MODEL=gpt-5.1

# Local LLM (Ollama, vLLM, llama.cpp, or other OpenAI-compatible API — used when LOCAL_AI=llm)
LOCAL_LLM_API_BASE=http://localhost:11434/v1
LOCAL_LLM_API_KEY=ollama
LOCAL_LLM_MODEL=llama3.3:latest

# Which local AI mode: 'ner-browser' (default), 'llm', or 'ner'
LOCAL_AI=ner-browser

# Set to 'false' to hide the Cloud AI option from the UI
CLOUD_AI=true

# Default UI language (optional): en, de, fr, es, ru, ar, zh
# DEFAULT_LOCALE=de

Azure AI Foundry setup

To use Azure OpenAI as the Cloud LLM, create an Azure AI Foundry resource and set:

CLOUD_LLM_API_BASE=https://YOUR-RESOURCE.openai.azure.com/openai/v1
CLOUD_LLM_API_KEY=your-azure-api-key
CLOUD_LLM_MODEL=gpt-5.1

The /openai/v1 path exposes an OpenAI-compatible API. No Azure-specific SDK configuration is needed.

Deployment profiles

Profile	CLOUD_AI	LOCAL_AI	Use case
Online demo	true (default)	ner-browser (default)	easyredact.io — Cloud AI + in-browser NLP
On-premise (GPU)	true	llm	Cloud AI + local LLM (Ollama, vLLM, llama.cpp)
On-premise (CPU)	true	ner	Cloud AI + spaCy NER for standard hardware
Air-gapped	false	llm or ner	No cloud connection at all

Docker

Pre-built images

Pre-built images are published to GitHub Container Registry on every release:

# Standard image (serves at /)
docker pull ghcr.io/datenlabor-bmz/easy-redact:latest
docker run -p 3000:3000 --env-file .env ghcr.io/datenlabor-bmz/easy-redact:latest

# Image with BASE_PATH=/easyredact (serves at /easyredact/)
docker pull ghcr.io/datenlabor-bmz/easy-redact-with-base-path:latest
docker run -p 3000:3000 --env-file .env ghcr.io/datenlabor-bmz/easy-redact-with-base-path:latest

Building from source

The Dockerfile bundles LibreOffice (DOCX conversion), Python + uv, and the German spaCy model (de_core_news_lg). By default it sets LOCAL_AI=ner:

docker build -t easy-redact .
docker run -p 3000:3000 --env-file .env easy-redact

For production deployment on linux/amd64 (e.g. when building on Apple Silicon):

docker buildx build --platform linux/amd64 -t easy-redact .

To use a local LLM instead of spaCy, override LOCAL_AI at runtime:

docker run -p 3000:3000 -e LOCAL_AI=llm --env-file .env easy-redact

DOCX upload and spaCy NLP are only available in the Docker build; they return HTTP 501 otherwise.

Subpath deployment

To serve the app under a subpath (e.g. acme.bund.de/easyredact/), pass BASE_PATH at build time:

docker build --platform linux/amd64 --build-arg BASE_PATH=/easyredact -t easy-redact .

This sets the Next.js basePath, which rewrites all routes, assets, and API endpoints. Configure nginx to forward requests without stripping the prefix:

location /easyredact/ {
    proxy_pass http://127.0.0.1:3000;
    proxy_set_header Host $host;
    proxy_set_header X-Real-IP $remote_addr;
    proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    proxy_set_header X-Forwarded-Proto $scheme;
}

Tech Stack

• Next.js 15 (App Router) + React 19
• MuPDF 1.27 — PDF rendering and redacted export via WASM + Comlink web worker
• OpenAI SDK — standard OpenAI client for both cloud and local LLMs; streaming chat completions with function calling
• spaCy (de_core_news_lg) — German NER, invoked via a Python script with uv run
• Tailwind CSS v4 + shadcn/ui (Radix primitives)
• IndexedDB (via idb) — client-side persistence for files, session state, and chat history

See Also

• datenlabor-bmz/redaction-ui — standalone React component library (@datenlabor-bmz/redaction-ui) for PDF viewing and redaction, published to npm for use in other applications
• datenlabor-bmz/redaction-rules — machine-readable FOI exemption rules by jurisdiction, fetched at runtime in FOI mode

License

AGPL-3.0. This project uses MuPDF which is licensed under the GNU Affero General Public License.

Credits

Built by the BMZ DataLab, the data science unit of Germany's Federal Ministry for Economic Cooperation and Development.

Funded by the European Union — NextGenerationEU.