Supported Formats

Every document belongs to a modality — text, tabular, image, audio, or metadata — which determines how it is analyzed and which redaction operators apply.

Formats

Text
Tabular
Image
Audio

Extension	Notes
`pdf`	Born-digital or scanned; scanned pages go through OCR
`docx`	Container format — embedded media is analyzed separately
`rtf`	Rich text
`txt`	Plain text
`log`	Plain text
`html`, `htm`	Markup
`xml`	Markup
`json`	Structured text

Extension	Notes
`csv`	Delimited rows
`xlsx`	Container format — sheets analyzed per part

Extension	Notes
`png`
`jpg`, `jpeg`
`tif`, `tiff`	Multi-page images supported

Extension	Notes
`wav`	Transcribed, then analyzed

Filtering by extension expands to the whole format’s extension set — passing jpg also matches jpeg, and tif also matches tiff.

Embedded metadata is its own modality. An EXIF tag or a document property can be surfaced as a redaction subject in its own right — such a finding carries a metadata event on its audit trail — so a photo’s GPS coordinates or a Word file’s author are findings like any other.

Container Formats

Formats like docx and xlsx are containers: they hold parts that may have their own modality. An image embedded in a Word document is analyzed as an image, in its own right. The detection’s analysis reflects this. The audit’s report carries a list of parts, each tagged by modality and identified by an id — a path of one or more segments naming where the part sits inside the document:

{
  "report": {
    "parts": [
      { "id": ["word/document.xml"], "modality": "text", "entities": [] },
      { "id": ["word/media/image1.png"], "modality": "image", "entities": [] }
    ]
  }
}

For a DOCX, a part’s id segment is its zip entry name. This is why one document can produce both text and image findings from a single detection, each redacted with operators appropriate to its modality.

OCR

Scanned documents have no selectable text layer and must be rendered to images before recognition. This is a codec setting, carried on the detection’s audit as codec.rasterMode:

{
  "codec": {
    "rasterMode": { "kind": "always", "dpi": 300 }
  }
}

Mode	Behavior
`auto`	Use the text layer where it exists, render only pages that lack one. The default
`always`	Always render pages to images at the given `dpi`, ignoring any text layer
`never`	Never render; rely on the text layer only, even if it is absent

always is for documents whose text layer is missing, garbled, or a watermark. Under auto, the text-layer-versus-scanned decision defers to the text path today; per-page rendering of scanned pages lands with OCR.

Filtering Documents

List endpoints filter by format, modality, name, or content hash:

const page = await client.documents.listDocuments(workspaceId, {
  formats: ["pdf", "docx"],
  modality: ["text"],
  search: "contract",
  limit: 50,
});

Every document carries the SHA-256 hash of its plaintext content, and hash is a list filter. Query it before uploading to check whether identical content already exists in the workspace.

Document Roles

Each document records the role it plays, which drives retention:

Kind	Description
`original`	Source document, uploaded or imported
`redacted`	Redacted output produced by a redaction

Machine byproducts — detection audits, review audits, and enrichment intermediates — are not documents. They are reached through their own endpoints and never appear in a document listing. A redacted output is reached from the redaction that produced it, through its outputDocumentId, so the original is never overwritten.

Overview

Features

Cloud

On-Premise

Supported Formats

Formats

Container Formats

OCR

Filtering Documents

Document Roles

Next Steps

Redaction Workflow

API Reference

​Formats

​Container Formats

​OCR

​Filtering Documents

​Document Roles

​Next Steps

Redaction Workflow

API Reference

Formats

Container Formats

OCR

Filtering Documents

Document Roles

Next Steps