Scanned PDFs and OCR

Who this page is for: the practice manager whose referrals arrive from a scanner or a fax gateway, and anyone asking what happens when a PDF contains pictures of words rather than words.

A referral PDF arrives in one of two forms. A digital PDF (exported from the referrer’s software) carries a text layer that is read directly. A scanned PDF (practice scanner, fax-to-email gateway, a photographed letter) is pages of images with no text layer at all - to a computer it is a picture. Refera detects which form each document is and recovers the text either way, so the extraction that fills the Inbox fields runs on every referral, not just the born-digital ones.

How detection works

Every PDF attachment is first checked for a usable text layer. If the layer yields real text, it is used verbatim and no OCR runs. If the document is image-only (or carries only the junk layer some scanner apps embed), each page is rasterised at 300 dpi and passed through optical character recognition. The recovered text then flows into the same rules-first extraction pipeline as typed text. OCR provenance is recorded with the result. Where a practice’s agreement and activation enable AI-assisted administrative extraction, Refera uses Amazon Bedrock to process recovered referral text, never images. Requests enter through the Sydney runtime endpoint and use Australian inference profiles (au.*), with inference in Sydney and Melbourne.

Rules run first, and the model is used only below the extraction confidence threshold. Haiku 4.5 is the default, with at most one Sonnet 4.6 attempt when Haiku’s result is invalid or a core field (patient name, date of birth, referrer or reason) is unresolved. Extraction returns administrative fields only. A model value is kept only if it appears in the letter. The Medicare number keeps its checksum validation. The model never sets urgency or category: any urgency wording is the referrer’s own, copied verbatim by fixed rules, and category comes only from a labelled field. Refera performs no clinical triage.

For that enabled path, referral text is processed over a private network path, with no internet route for the extraction function. Bedrock data retention is set to “none” in both Australian regions, invocation logging is off, and customer content is not used for training.

Confidence and manual review - never silent

OCR is honest about its own limits. Every recovered word carries a confidence score, and the document’s mean confidence travels with the extraction:

• At or above the review threshold - fields are extracted and shown beside the available source excerpt and attachment metadata.
• Below the threshold (a grainy fax, a skewed photocopy, handwriting) - the document is flagged for manual review. The recovered text and any fields it produced are still shown, marked unverified, with the available source excerpt and attachment metadata beside them. Nothing is silently dropped or guessed at.
• Unreadable or corrupt attachments are flagged the same way - one bad attachment never sinks the rest of the message.

Refera does not yet claim a deployed original-PDF archive. The practice keeps the original in its existing mailbox or document system while Refera’s connection checks use only patient-free test files. Real-data activation remains blocked until retention, authenticated retrieval, restore/readback and complete source export are proven end to end.

Patient-free test lane and launch checks

Two lanes, stated plainly:

• Patient-free setup: OCR runs locally on the capture host using open-source tooling (poppler-utils and Tesseract). The OCR step recovers text locally. Capture-off setup accepts clearly tagged, patient-free test documents only.
• Real-data activation: source-document and practice activation checks must pass before patient documents are accepted. Where the practice’s agreement and activation enable administrative extraction, only recovered text can reach the Australian Bedrock path described above. Images and PDFs are not sent to Bedrock by this feature.

What OCR does not do

OCR recovers text - that is the whole job. It does not interpret the referral, does not rank it, and does not compute anything from the clinical content. Referrals remain in arrival order, and any urgency wording is captured only as the referrer’s verbatim words. Refera tracks referral admin only. It does not triage patients.

Related

• Fax to email - where most scanned PDFs come from.
• Upload - paper and scans - the planned path for the paper pile, not currently available in the connected Inbox.
• Microsoft 365 mailbox - the mailbox the scans land in.