Detect
Deterministic rules run first. When enabled, an LLM reviews the partially redacted text for contextual or indirect identifiers, and every proposed action is validated before application.
Semred combines context-aware redaction with an inspectable safety protocol—from the first rule pass to the final export. Remove PHI, PII, and confidential detail while preserving meaning.
Rules run on your host · External AI off by default · Leak-checked exports
The reason teams cannot use a general‑purpose model is the reason Semred exists. Four guarantees hold on every run. See the full protocol
Redaction exists because the document cannot be sent anywhere. So Semred draws the boundary first and works inside it. Watch one document make the trip.
Your host · nothing leaves without the flag
Deterministic detection runs before anything else and never makes a network call.
The semantic pass only runs when the deployment sets SEMRED_ALLOW_EXTERNAL_LLM=true. Each run records whether it did.
Stage two receives the stage‑one output, with names, identifiers and dates already replaced.
No source text, prompts, model output or replacement mappings are written to service logs.
TXT, DOCX and PDF exports are re‑extracted and compared against the records before they download. Residual data blocks the export.
Policy fingerprint, detector and prompt versions, model name and validation state ship with each result.
PDF export writes a new text PDF. It does not scrub hidden layers, metadata or embedded images of an original PDF.
Deployment decisions that stay with you, such as identity provider, encrypted storage and retention, are listed in the security runbook we share with design partners. Discuss a deployment
Semred separates deterministic coverage from contextual review. Each applied change is inspectable, and the run records which stages actually completed.
Deterministic rules run first. When enabled, an LLM reviews the partially redacted text for contextual or indirect identifiers, and every proposed action is validated before application.
Overlapping and duplicate spans are normalized within each stage, with stable labels and conservative conflict handling before that stage rewrites text.
Accepted spans are replaced locally with stable placeholders. Unchanged source text is preserved rather than silently regenerated.
A similarity score compares the source and result. Embeddings are used when enabled and available; otherwise a deterministic lexical fallback is used. It is not a safety or coverage score.
Every applied replacement records its type, source value, replacement, stage, and review metadata. Audit trails can be exported as JSON or CSV.
A pattern matcher sees nine digits. Semred sees whether those digits are an SSN, a case number, or a quantity — and whether the sentence still works once they’re gone.
Talk to us about your corpusReplacement records · inspector view
Redaction that destroys meaning deserves review. Semred compares the source and result so you can spot substantial change; the score is a similarity signal, not proof of safe coverage.
What the score does not meanIllustrative sample. Similarity describes meaning, not coverage.
Each replacement carries its type, original value, substitute, and reason. Export the whole trail as JSON or CSV when compliance asks how a document got that way.
Where the data goes, and where it never doesAudit records · JSON and CSV export
Semred is designed for environments where a single exposed record is unacceptable.
De‑identify clinical notes, referrals, and messages while preserving the clinical detail that makes them useful downstream.
Sanitize contracts, discovery documents, and briefs before they reach vendors, co‑counsel, or a model.
Strip internal docs, tickets, and logs of identifying detail before they leave your environment.
Redaction tools that promise completeness are the ones that get people hurt. Here is where our guarantees stop.
Semantic similarity describes how much meaning survived. It says nothing about whether every sensitive span was found.
An empty replacement list means nothing matched — not that there was nothing to match. Review still applies.
Semred is built to make human review fast and evidence‑backed, not to remove it. Spans the model alone proposed are flagged as such.
That’s the case we built this for. Tell us what you’re sitting on and we’ll show you what comes out the other side.
Request accessWe work directly with teams that have real redaction problems — not toy demos. No self‑serve tier while we’re shaping the product around design partners.
Work directly with us to shape Semred around your workflows.
Share a bit about your use case and we’ll get in touch.