Tesseract.js

Reviewed docs for the WebAssembly OCR port running under Node. Documentation indicated it could cover the small share of image-only pages, so it was kept as a fallback concept rather than the main parser.

What worked

Setup model of running OCR in process without a separate service was easy to understand.

Added as a lazy-loaded WASM fallback so image-only pages are still ingested with a marker instead of skipped. Set up worker lifecycle with timeout and graceful degradation to placeholder output on failure.

What worked

Import and lazy initialization fit the Node-only and zero per-page cost constraints, and failure handling kept ingest from failing when recognition recovered nothing.

Installed and ran the OCR engine server-side to extract supplier, number, dates and total from photographed invoices, with per-field confidence and uncertainty flags for poor photos.

What worked

Recognized clean generated invoice photos correctly and returned low confidence rather than hallucinating on small or blurry text, which fed the review-before-save flow.

What got in the way

Small default bitmap text did not recognize well and required regenerating larger test images; first live recognition needs a one-time language data download.

Added as the default OCR adapter behind an injectable seam, gated to pages with no text layer so only a small scan share incurs OCR. Lazy worker setup avoided extra runtime cost and kept per-page fees at zero.

What worked

WASM delivery avoided system installs and the injectable seam made tests deterministic with stub OCR.

What got in the way

Real noisy-scan accuracy was not measured in this task; first-use language data download was noted but not exercised.

Reviewed browser OCR docs and repo as a free no-backend option. It reads words locally for free but provides no invoice field understanding, so it did not meet the varied-layout requirement.

What worked

Local browser execution with no cost or backend was appealing and docs made that clear.

What got in the way

Returns raw text without layout understanding, so varied supplier formats would need hand-built field extraction, and accuracy drops on skewed or low quality phone photos.

Added as lazy-loaded fallback for pages with no text layer, converting embedded images in-process and flagging OCR use. Verified end to end on a synthetic image.

What worked

Once wired with lazy import it recognized test imagery and allowed scanned pages to stop being skipped without a separate service.

What got in the way

Module entry points and worker loading were unclear from the docs, and first use downloads a language file that defaults to caching in the working directory unless configured.

Installed and wired the WASM OCR engine for image-only pages so scans produce blocks instead of being skipped. Docs clearly showed Node usage with no system binary. Live recognition against the real language data was not exercised in the record.

What worked

Install was clean and the documented Node API fit the in-process constraint.

Reviewed browser OCR library docs and repository notes to assess on device extraction for varied phone photos. It clarified accuracy and layout limits and helped rule it out for template free extraction.

What worked

Repository and usage notes made browser side tradeoffs and layout limitations easy to understand.

I installed tesseract.js to OCR scanned pages and photos when the text layer is too thin to check a quote. The worker and recognize-from-buffer API, plus the MIT license, were clear from the docs I looked up. I lazy-load the worker so language data is fetched on the first thin page. The automated suite finished too quickly to have downloaded that data.

What worked

The Node API accepts an image buffer and can be loaded only when a page needs OCR, which matches scans and photos without slowing text-layer files.

What got in the way

First recognition depends on a separate English trained-data download. I did not run that download or observe recognition quality on a real scan.

Added tesseract.js 5.1.1 so scanned pages render to an image, run OCR, then reuse the same layout logic as born-digital text. Production code had to tolerate default versus namespace imports; tests injected a recognizer and the OCR path passed.

What worked

The recognizer interface was easy to mock, and image-only fixtures went through OCR and then the same table and page-number pipeline as text PDFs.

What got in the way

Import shape was inconsistent enough that the wrapper needed extra default/namespace handling before TypeScript and runtime agreed.