Piixie Docs

Piixie creates anonymized copies of documents. The original file stays in place, and the output copy is written to the local Piixie data directory for review and export.

Open Piixie and let the local model download on first run.
Drop a document or choose it from the file picker.
Select redaction, replacement, or synthetic mode.
Review the anonymized copy and replacements table.
Send the safe copy to an LLM, ticket, test fixture, or collaborator.

Anonymization modes

Redaction

PII is replaced with visible redaction markers. Use it when the safest output is a document that reveals no sensitive value shape.

Replacement

PII becomes stable tokens such as name and email placeholders. Use it when consistency matters during review.

Synthetic

PII becomes plausible fake values generated locally through Piixie's bundled Faker bridge. Use it for demos, training samples, tests, and prompts that need natural context.

Supported files

Current desktop support is focused on text files, Markdown, DOCX, PDFs, and data-table workflows. The roadmap is to support every file type that needs anonymization without adding CGO dependencies.

OCR is not a separate requirement for image-backed documents. Piixie's local model is vision-capable, so the anonymization pipeline can inspect rendered pages and image regions directly instead of depending on a brittle OCR preprocessing step.

Why LLMs for PII detection

PII is not always a clean token shaped like an email address, credit card number, or phone number. In real documents, private information is often implied by relationships between fields, surrounding paragraphs, tables, captions, images, and repeated references. Piixie uses a local LLM because those relationships are exactly where rule-based detection starts to break down.

A rules engine can match john@example.com. It cannot reliably decide that "the claimant", "her dependent", "the upstairs tenant", and an account row three pages later all refer to the same private person. An LLM can use context to connect those references, classify the entity, choose the right anonymization mode, and preserve consistency in the rewritten document.

Read the detection guide

Automation

Piixie is designed to run as a server and as a CLI so privacy gates can live inside existing workflows. See the automation guide for example batch commands and deployment patterns.

Open automation guide

External LLMs

The default posture is local anonymization first. Teams can still connect preferred external providers after the raw document has been transformed into a safer copy.

sequenceDiagram
  participant User
  participant Piixie
  participant LocalModel as Local model
  participant Provider as Optional external LLM
  User->>Piixie: Open raw document
  Piixie->>LocalModel: Detect and transform PII
  LocalModel-->>Piixie: Anonymized content
  Piixie-->>User: Safe copy
  User->>Provider: Send safe copy when needed

Privacy model

Piixie is local-first, not cloud-first. The architecture is built around local extraction, local model execution, local output review, and optional external provider use only after anonymization.