Does Karchu always run OCR?

No. A PDF with a usable text layer is read directly, which is faster and more accurate. OCR runs only when the text layer is missing or unusable, for example on a scan or a photograph.

What happens to a document Karchu cannot read?

It is marked failed with a specific reason, such as a password protected PDF, an empty page range or an image too low in resolution. The original file is kept and processing can be retried without uploading again.

How many pages can one document have?

Up to 300 pages. Every page inside that limit is processed.

Document scanning

Pipeline

upload
  -> validate (type, size, pages, readability)
  -> classify (statement, card statement, receipt, invoice, bill, expense report)
  -> extract text (PDF text layer first)
  -> OCR fallback (only when text is missing or unusable)
  -> parse transactions
  -> validate transactions
  -> categorize
  -> reconcile
  -> complete or requires_review

Karchu accepts a financial document, validates it, works out what kind of document it is, and extracts text before any parsing happens.

What this stage does

Validate the file type, size and page count before processing and return a specific error when a file cannot be read
Classify the document as a bank statement, credit card statement, receipt, invoice, bill, expense report or unknown, with a confidence score
Ask the user to confirm the document type when classification confidence is below 0.7
Read the PDF text layer first and only fall back to OCR when the text layer is missing or unusable
Preprocess photographed and scanned images: deskew, orientation correction, grayscale and contrast normalisation
Keep the uploaded original alongside the processed representation
Process every page of a multi page document, up to 300 pages
Record the processing state of each document: uploaded, validating, validated, extracting, ocr_processing, classifying, parsing, validating_transactions, requires_review, completed, failed

Accepted files

PDF (digital text layer)
PDF (scanned, image only)
PDF (multi page, up to 300 pages)
PNG
JPG and JPEG
HEIC where the browser can decode it
CSV
XLS and XLSX
ODS
TXT

Limits

Password protected PDFs are rejected with an explanatory error rather than guessed at
A photograph too blurred for OCR returns a low confidence result flagged for review, not invented text
HEIC support depends on the browser being able to decode the image

Processing states

Every document carries a state you can see in the app: uploaded, validating, validated, extracting, ocr_processing, classifying, parsing, validating_transactions, requires_review, completed, failed. Each state change is recorded with a timestamp, so an administrator can tell exactly where a document stopped and why.

Retries

A failed document can be reprocessed from the stored original. Attempt count, error and stage are tracked, and there is a hard limit of five server side attempts so a broken file cannot loop.