1
Ingest
PDFs, photos, scans, zips. Text is read natively where it exists and OCR'd where it doesn't. Duplicates are caught by content hash.
Documents in.
Structured data out.
Extractly reads invoices, bank statements, lab reports and hospital paperwork — PDFs or photographs — works out what each one is, pulls every field into your database, and shows you exactly which page it came from.
Private preview · the app asks for credentials
Source · page 1
ACME SUPPLY CO.INV-2291
Billed toNorthwind Ltd
Date issued14 Mar 2026
Subtotal1,240.00
Tax (8.25%)102.30
Total due1,342.30
Extracted
invoice_noINV-2291p10.98
vendorAcme Supply Co.p10.96
issue_date2026-03-14p10.94
subtotal1240.00p10.97
tax102.30p10.88
total1342.30p10.97
subtotal + tax = total ✓
The pipeline
Five steps, and you stay in the loop for the one that matters.
Everything that can be decided without a language model, is. The AI is called when the document is genuinely ambiguous — not on every page.
2
Classify
Keyword and pattern signals name the document type on their own. Only a weak or contested score escalates to the model.
3
Extract
The matching template tells the model exactly which fields to find. Money, dates and totals are parsed into real types, not strings.
4
Review
Source on the left, editable fields on the right. Every value carries a confidence score and the page it was lifted from.
5
Export
One document or every document of a type — as CSV, as a spreadsheet, or pushed straight into your own API or database.
Templates
Seven document types, and adding an eighth is a JSON file.
Templates are data, not code. Each one names its fields, their types, and the arithmetic that has to hold true.
Invoicesubtotal + tax = total
Receiptline items + tender
Bank statementopening + Σ = closing
Lab reportanalytes + ranges
W-2wages & withholding
EOBplan paid + resp = allowed
Medical recordencounter + diagnoses
Need purchase orders, remittance advice, or a shipping manifest?
Drop a template.json into the registry — no rebuild, no deploy.
Why it's trustworthy
An extraction you can't check isn't worth having.
Anything that lands in your database can be traced back to the pixel it came from.
Page-level provenance
Every field remembers its page and carries a confidence score. Low-confidence values are flagged for a human before anything is exported.
The maths has to hold
Invoices must balance. Bank statements must reconcile from opening balance through every transaction to closing. When they don't, Extractly says so instead of quietly filing a wrong number.
Map once, push forever
Point total at amount_cents, apply a transform, preview the rows, then push to a REST endpoint, a database table, or a sheet. Dry-run first, always.
Sensitive pages never leave
Vision runs on Qwen-VL in production and OpenAI locally — but documents you mark as sensitive are read by an on-box OCR engine instead, so no PHI crosses the wire.
Connected apps
Right-click a file in StorageSpace. Send it here.
Extractly plugs into your StorageSpace drive with a signed hand-off — the file travels with a short-lived token, gets processed, and appears in your review queue. No download, no re-upload, no copy sitting on a laptop.
StorageSpaceyour drive
Signed tokenexpires fast
Extractlyreview queue
Stop typing numbers off a screen.
Extractly is live and processing documents right now. Sign in to upload your first one.
Private preview — the app asks for credentials