Documentation / examples/demos/ocr-document-review/README.md

Evidence-first OCR document review demo

This is the canonical reusable asset set for the PixieCore OCR story. It uses one scanned PDF, one small Extractor Blueprint, one versioned dataset case, and one typed result in video, article, and conference formats.

Run the offline contract rehearsal after building PixieCore:

npm run build
node examples/demos/ocr-document-review/run.mjs --mode=mock

mock runs the real PixieCore validation and attachment pipeline with the dataset's declared expected output. It proves the contract and presentation flow without a credential. It does not measure OCR model accuracy.

Use --mode=real only with an explicitly configured provider/model that supports the scanned PDF input. --mode=both prints fixture and real-provider results side by side. Do not publish a quality claim from one run; use the repeated benchmark and scorecard process in docs/blueprints/evaluation.md.

The selected case extracts Riley Chen and Chicago, USA, returns the absent amount as not_found, and emits an ocr_required warning rather than inventing a value. The application boundary may next map that typed result into a form, human-review step, or workflow. This demo itself performs no persistence, submission, approval, or notification.

See demo.yaml for the canonical assets, commands, story order, and evidence labels. See docs/blueprints/ocr-demo.md for the video shot list, article outline, and conference runbook.