document-extractionlisted

# Document extraction Turning a document into a row is easy. Turning it into a row you can *post to a ledger without a human looking* is the hard part, and it is entirely about knowing which fields you are unsure of. ## When to use this Reading invoices, receipts, purchase orders, contracts, packing slips or forms into structured data. Single file or a directory. ## How to run it 1. **Get text out of the file first.** Decide which kind of document you have: ```bash python3 scripts/extract.py detect invoice.pdf ``` It reports `text` (a digital PDF with a real text layer), `scan` (images only, needs OCR), or `mixed`. 2. **Extract.** ```bash python3 scripts/extract.py run invoice.pdf --schema invoice --json ``` For a scan, the script shells out to `pdftotext`/`tesseract` when they are installed and tells you plainly when they are not. Do not silently return empty fields. 3. **Read the confidence block before the data block.** Anything below the threshold is a field a human must confirm. Say so in your answer, by name. 4. **Never invent a missing field.** A null with a reason beats a plausible guess, every time. This is the rule the whole skill exists for. ## The schema `references/schemas.md` holds the field sets. The invoice one, which covers most work: | Field | Type | Notes | |-------|------|-------| | `invoice_number` | string | As printed. Do not normalise away prefixes | | `invoice_date` | ISO date | Watch DD/MM vs MM/DD —