Sensitive data,
truly gone.

What it is

Redaktion is an automated detect-and-remove engine for sensitive documents. It finds personal and corporate information across 45 configurable categories, burns opaque boxes at exact coordinates with true content removal, strips document metadata, and produces a forensically inert artifact — then verifies its own work before calling the job done.

It runs as a set of workers behind a queue. You hand it a document and a job notification; it hands back a redacted artifact, a structured summary, and a per-box audit trail that never contains the sensitive text itself.

How you integrate

Redaktion is a queue-driven worker, not a REST endpoint. That is deliberate: redaction is a long, bursty, retry-heavy workload, and a durable queue plus a job record is a better contract for it than an HTTP request that can time out halfway through a 300-page scan.

Four steps, and the shape does not change with volume:

Queues

Every queue is durable and dead-lettered — a matching dlq- queue receives anything that cannot be processed, so a poison message is quarantined rather than looping forever. Queues must be declared with the dead-letter arguments the workers expect; a mismatched declaration is rejected outright rather than silently accepted.

A PDF whose text layer is too thin is re-routed to the OCR pipeline automatically, and a very large scan to the extra-large lane — without consuming retry budget, because a handoff is not a failure.

Job contract

What you publish:

{
  "type": "redaction",
  "document_id": "0f1e2d3c-…",
  "case_id": null,
  "client_id": "acme",
  "file_name": "contract.pdf",
  "file_path": "tenant/<uuid>/contract.pdf",
  "file_extension": "pdf",
  "redaction_type": "pdf",
  "timestamp": "2026-08-03T09:12:44Z",
  "storage_backend": "minio"
}

Workers consume with a prefetch of one, so a single worker takes one document at a time and a slow job never blocks a queue behind it.

Result & progress events

Rather than polling the record, you can subscribe. Both event types are versioned and field-locked across every worker implementation.

Result

{
  "type": "redaction_result",
  "timestamp": "2026-08-03T09:14:01Z",
  "document_id": "0f1e2d3c-…",
  "client_id": "acme",
  "attempt": 1,
  "outcome": "completed",
  "status": "completed",
  "redacted_file_path": "tenant/<uuid>/contract_redacted.pdf",
  "redacted_file_size": 481233,
  "redaction_summary": "{…}",
  "redaction_has_warnings": false,
  "redaction_vision_qa_verdict": "pass",
  "redaction_quality_verdict": "ok",
  "redaction_quality_issues": [],
  "redaction_rules_version_applied": 7,
  "audit_rows": [ … ]
}

Progress

{ "type": "redaction_progress", "timestamp": "…", "document_id": "…",
  "attempt": 1, "progress_page": 7, "processing_stage": "detect" }

Stages are named and ordered, so a progress bar is a lookup rather than a guess: preflight → detect → apply_and_verify → upload_and_complete on the PDF path, and ocr_preflight → ocr_pipeline → ocr_upload_and_complete on the OCR path. Heartbeat-only events omit the page and stage.

Workers also expose /health and Prometheus /metrics for your own monitoring.

What it detects

Forty-five rule keys, each individually enabled or disabled per tenant.

How toggles behave

Detected values are also normalised through an alias map, so a model that reports DOB, iban or Driver License lands on the right canonical key rather than being lost.

Regional ID packs

Generic PII tools do badly on identity documents outside the US. Redaktion ships country packs where the pattern is paired with the checksum that proves it — and a value that fails its checksum is rejected rather than redacted, so look-alike numbers don't produce noise.

Machine-readable zones are always removed — an MRZ encodes every field on the page, so leaving it visible leaks the document you just redacted.

A further set of label-anchored patterns catches values that only a nearby label identifies. Those redact the value and not the label, so a form stays readable as a form.

Detection stack

Five layers, run in a fixed order, merged once with a defined precedence — deterministic matches beat statistical ones, and statistical beat generative.

The recognition layer runs remotely and is fail-soft: if it is unavailable, detection degrades to the deterministic layers and a circuit breaker stops the worker hammering a service that is down. A downed component slows detection; it does not take redaction offline.

Document-type skills

Fixed-layout identity documents get dedicated handling — Aadhaar, Emirates ID, and any ICAO-standard passport — that knows where the face, signature, number and MRZ sit rather than searching for them. A skill only takes over when it recognises the document confidently; otherwise the generic pipeline runs. Every skill failure path falls through rather than failing the job.

The verification pass

This is the part that separates Redaktion from a tool that draws rectangles.

After boxes are applied, the engine re-reads its own output — OCRs the redacted file and runs the same detection pipeline over what it can still see. Anything found there is a leak by definition. Because it reuses the same rules, its false-positive behaviour is identical to the primary pass, rather than a second, noisier detector second-guessing the first.

On the PDF path there is a further chain: a fast text-layer check, then an OCR check, then any residual is re-boxed and re-verified, with page rasterisation as the last resort. A vision QA pass then asks a model a deliberately narrow question — we drew N boxes, do you see N boxes, and is anything still legible? — because that is far more reliable than asking a model to find all PII from scratch.

Geometry is audited too: a job where boxes were drawn but landed over the wrong tokens is flagged, which is a failure mode a simple count check would pass.

Statuses

Escalation is triggered by concrete signals, not a vague score: uncovered PII or low confidence from vision QA, partial OCR coverage, a degraded OCR engine, too few boxes for the entities found, boxes landing in the wrong place, a photo-mode document, or a low-quality scan where nothing was found at all.

Configuration

Tenant settings, applied per job.

Rule toggles take effect on the next job — the enabled set is re-read before each document, with no worker restart. A transient failure to load keeps the last good set rather than clearing it.

Supported inputs

Arabic is handled properly rather than nominally: right-to-left reordering, Arabic-Indic and Devanagari digit normalisation before matching — while the original characters are what get located on the page — plus Arabic address and road patterns.

Output guarantees

Removal is real

On PDFs the engine applies true redaction annotations that delete the underlying content, then scrubs metadata, embedded attachments and page remnants, and saves with full garbage collection. The result cannot be un-redacted by copying a layer, selecting invisible text, or reading document properties. On images the pixels are overwritten, not covered.

Box geometry is padded for serifs and shrunk slightly so bars don't bleed into adjacent rows, overlapping boxes are collapsed, and a form label is never drawn over.

Fail-closed

A page the pipeline cannot process confidently ships fully blacked out and the document is flagged. The safe failure is a page you re-run, not a page that leaks.

The audit trail never holds the data

One row per drawn box:

The digest is keyed per document. That preserves duplicate analytics inside one file while making it impossible to correlate "this identifier appears in documents A, B and C" across your corpus — which an unkeyed hash would allow to anyone holding the table. Audit writes are best-effort by design: a failing audit insert never fails the redaction that produced it.

Encryption

Sources are read through an AES-256-GCM envelope and the redacted artifact is written back to the same backend, re-encrypted. The summary attached to each result reports counts and categories with fingerprints rather than examples, so a summary is safe to log.

Failure & retries

Failures are typed, not generic — you can route on them. Hard failures include model timeouts and rate limits, OCR timeouts, empty text extraction, storage read/write failures, verification failures, unsupported input, files that are too large, and input quality below threshold. Warnings — partial chunk failures, suspicious zero-entity results, low-confidence vision QA, partial OCR coverage, misplaced boxes — surface on the result rather than failing the job, and escalate to review where they matter.

Your data

Source documents stay encrypted at rest under a key you control, and redacted artifacts are written back encrypted. Removal is real, so a redacted file cannot be reversed into the original. The audit trail records where and why each redaction happened without ever storing the sensitive text.

Sensitive content is necessarily processed in memory to be detected, and detection uses model services to adjudicate ambiguous spans. Working files are swept on a short timer. Processing runs on EU/Swiss-resident infrastructure. Retention, subprocessors and erasure are set out in the privacy policy.

Availability & access

Redaktion is a SeKondBrain Labs product. The engine runs in production inside our legal platform today. As a standalone service it is at early access: we integrate it with your pipeline and provision the queues, storage and credentials with you, rather than offering an open self-serve endpoint. It is licensed, not open source.

Where it runs

What you get when you're set up