Document Extraction (Textract)

Asynchronous document text extraction powered by AWS Textract.

Textract extraction is an alpha capability. It is gated behind the textract_extraction feature flag and the confidentiality action-gating framework, so it is not available by default.

Introduced in v0.12.0, the Textract pipeline extracts text from uploaded documents asynchronously. Access is controlled at two layers: a per-user feature flag and the project confidentiality policy.

Overview

Asynchronous extraction: documents are queued and processed in the background
Per-document status badges reflect the current extraction state
Extracted text feeds the existing document and search subsystems

Access requirements

A user can trigger extraction only when both conditions hold.

The user holds the textract_extraction feature flag.
The project confidentiality policy permits extraction for the target project (see below).

Extraction status

Each document surfaces a status badge indicating where it is in the pipeline. Status is shown per document in the documents view so users can see at a glance which documents have been processed.

Confidentiality gating

Extraction is a gated action under Confidentiality Action Gating.

For lower-severity projects, extraction is allowed.
For medium-severity projects (C.2 and C.3), extraction is gated: it returns warn_and_override, and proceeding requires an explicit override that is written to the audit trail.
Higher-severity classifications block external processing at the AI-routing layer.

Because Textract is an external AWS service, extraction on classified projects is governed by the same confidentiality principle that blocks external model providers.