Skip to content

Importing PDFs

Ingest turns a source report into a stored report, structured rows, and review diagnostics.

Import a report

Open [Ingest], then drop one or more PDF files onto the import area or use the file picker. Batch imports are processed sequentially so each file has a clear result and progress state.

Supported report sources

bloody-level supports Portuguese (PT-PT) pathology and laboratory PDF reports from CUF and Germano de Sousa. Text-based PDFs are handled by the local PDFium extractor. Scanned or text-poor reports can use the locally compiled OCR tiers when it is configured and available.

Provider layouts may vary by department, report type, and revision. After every import, compare the patient, date, analyte names, values, units, flags, and reference ranges with the original PDF. CUF and Germano de Sousa are document sources referenced for compatibility; bloody-level is independent and is not affiliated with either provider.

The pipeline records these stages:

  1. Verify that the file exists and is a PDF.
  2. Hash the source bytes with SHA-256.
  3. Extract text with PDFium.
  4. Use OCR when the extracted text is sparse and the selected tier is available.
  5. Parse and normalize rows against the local analyte Library.
  6. Encrypt the managed source copy, then write the report, rows, and audit events to the encrypted vault.

Duplicate handling

The source SHA-256 is the identity check for an imported file. Importing the same bytes again is treated as a duplicate rather than creating a second report. A report's original PDF is copied into application data so later review does not depend on the original path continuing to exist. The managed copy is encrypted at rest; opening it creates a temporary hand-off for the system PDF viewer.

Ingestion tiers

Settings presents three independent ingestion stages:

  1. Tier 1 — PDF extraction: the local PDFium baseline. This must be enabled for imports to proceed.
  2. Tier 2 — OCR: Tesseract fallback for sparse-text PDFs. It is used only when the toggle is enabled and the native executable and language data are available.
  3. Tier 3 — Hybrid OCR + LLM: when Tier 2 reports low OCR confidence and the Phi-4 model is enabled, the app loads it on demand, repairs the OCR text, and parses that repaired text. The stored report is marked as Tier 3 only when a non-empty model output was actually used. If the model is missing, unloaded, or fails, the app keeps the Tier 1/2 text and records the reason. olmOCR-2 remains an optional vision resource and is shown as unavailable until its native inference runtime is ready.

Distributed builds compile the optional Tesseract, olmOCR-2, and Phi-4 integrations, but compiled code is not the same as an enabled, downloaded, or loaded runtime asset. Choose the stage toggles and model paths in Settings. The UI shows availability, download progress, cancellation, and diagnostics instead of pretending a missing backend ran. A normal text-based report should remain Tier 1: enabling a higher tier makes it an escalation fallback, not a forced second pass over every document.

After import

Open the report detail and verify the patient, date, analyte names, values, units, flags, and reference ranges against the original PDF. Treat unmatched, unparsed, or low-confidence rows as review work. See reviewing results for the available actions.

When an import fails

The failure is kept with the import attempt where possible. Read the diagnostic message, check troubleshooting, and retry after correcting the source file or local backend. A failed parse should not be treated as an empty normal report.