tika-ocr

Custom Apache Tika image for tika.populate.tools with OCR enabled for Spanish, Catalan and Basque.

What's inside

• Base: apache/tika:latest-full (Tika + Tesseract + English).
• Adds: tesseract-ocr-spa, tesseract-ocr-cat, tesseract-ocr-eus.
• Loads tika-config.xml so the PDF parser uses ocrStrategy=auto — OCR runs only when a page has no text layer, so text-PDF extraction stays fast.

Timeouts

• Per-task ceiling is 30 minutes (taskTimeoutMillis=1800000).
• Tesseract per-OCR-call ceiling is 30 minutes (timeoutSeconds=1800).
• Clients can shorten per-request via X-Tika-Timeout-Millis and X-Tika-OCRtimeoutSeconds. They cannot raise above these defaults.

Image

ghcr.io/populatetools/tika-ocr:latest

GitHub Actions rebuilds the image:

• on every push to main,
• on every v* tag push (versioned releases),
• nightly at 04:00 UTC, so the upstream apache/tika:latest-full security updates get picked up,
• on demand via workflow_dispatch.

Tags published:

• latest — rolling, follows main.
• vX.Y.Z, X.Y, X — published when a vX.Y.Z git tag is pushed.
• sha-<short> — every build.
• nightly — the scheduled rebuild.

Local smoke test

docker build -t tika-ocr:dev .
docker run --rm -p 9998:9998 tika-ocr:dev

curl -sL -o /tmp/anexos.pdf \
  "https://contratos-files.gobierto.es/documents/tenders/6a89c71d8756f85720bb40f2c631b13d/Anexos.pdf"

curl -sX PUT --data-binary @/tmp/anexos.pdf \
  -H "Content-Type: application/pdf" \
  -H "Accept: text/plain" \
  http://localhost:9998/tika | wc -c

A successful run prints a body well over 20 bytes containing recognisable Spanish text. The same request against vanilla apache/tika:latest-full returns ~20 bytes of newlines for image-only PDFs.

Deploy

Pull the published image and run it however you deploy containers:

docker pull ghcr.io/populatetools/tika-ocr:latest
docker run -d -p 9998:9998 ghcr.io/populatetools/tika-ocr:latest

Why this exists

apache/tika:latest-full ships Tesseract + English only. Image-only PDFs from Spanish public administrations (scanner output, Hewlett-Packard MFP, etc.) silently extracted to empty strings, masking ~22% of indexed documents in contratos.gobierto.es. See the parent investigation in the contratos repo.