Native text comes first

Readable embedded PDF text and supported Office/text content can be extracted without OCR. The OCR worker uses PDF.js for embedded PDF text; secure intake provides the canonical extraction for supported Office and text formats. Native extraction reads existing text rather than recognising characters from an image.

Scanned/image-only pages require an enabled OCR path. A mixed PDF can retain native text on readable pages and OCR image-only pages. Tesseract does not render an empty Office document; conversion or a supported alternative is required.

Choose when OCR runs

OCR behaviour and OCR engine answer different questions. When readable text is missing runs OCR only where the text path needs it; Always re-run OCR explicitly requests OCR; Never run OCR prevents OCR regardless of an engine or escalation preference.

Choose the permitted engine in processing options/profile controls. Deployment default uses the configured environment. Local OCR uses Tesseract only. AWS OCR explicitly selects Textract where permitted. Automatic tries local OCR first; paid AWS escalation requires both deployment permission and the saved job/profile opt-in.

Configure and test

  1. Open the relevant processing profile or document-stage settings. Review OCR behaviour independently from the engine policy.
  2. Select an installed language code, minimum confidence and permitted local page limit where the controls are available. Language packs must exist in the worker image.
  3. If Automatic is selected, decide whether AWS escalation is permitted for this workload. A checkbox cannot override a deployment that prohibits AWS.
  4. Save the profile and test representative native PDFs, scans, mixed PDFs and difficult documents before using it for a customer workload.
  5. Inspect the actual provider, page coverage, quality flags and source evidence on the result. Resolve low confidence, truncated text or missing pages through review.
Local product example · sample dataClick to enlarge ↗
Set the OCR engine and quality limits

This example selects local Tesseract OCR with English language codes, an 80% confidence threshold and a 25-page local limit. These are illustrative settings, not recommended defaults for every workload. AWS escalation is unchecked.

What OCR does not prove

Tesseract provides recognised text and word geometry for supported PDF, PNG, JPEG and TIFF inputs within configured resource limits. It is not equivalent to every Textract forms, tables or signature-analysis capability.

OCR does not verify identity, validate an amount or approve a document. Financial/date/identifier checks and human-review gates still apply. Low-quality or incomplete local OCR requires review even if a permissive review preference was selected. Local OCR execution and customer metering are separate: it does not automatically make processing unmetered.

Investigate incomplete recognition

  1. Compare page coverage with the original, including rotated, clipped or low-resolution pages.
  2. Check actual provider, installed language and resource/page limits.
  3. Verify critical values against source regions and correct through review.
  4. Test changed preparation/profile settings on samples before rerunning live batches.