How it works
From a folder of PDFs to a clean, checked table
You describe the data once and check a few examples. PDFSifter’s AI learns the rest, reads every paper, and shows you exactly which values need a second look.
Step 1
Describe the data you need
A template is one kind of PDF and the values you want from it. Write a short brief in plain English, then list the fields. Each field has a type, so numbers stay numbers and dates stay dates.
Step 2
Add three examples
A training pair is a PDF plus the answer you expect. Choose “Label in PDFSifter” and the AI drafts the values for you to check against the paper, or upload a CSV or JSON file with one row per record.
- Keep values exactly as the paper prints them
- Leave a cell empty when the paper doesn’t report it
- Pick papers that look different from each other
Step 3
Train the AI, and see how well it did
Press Train. The AI writes the extraction instructions itself and PDFSifter measures them against your pairs, so the accuracy you see is how often it matched your own answers.
Every run is kept in History with its accuracy and cost. If a retrain on the same pairs scores lower, you keep the version you had.
Step 4
Run it on a whole folder
Choose a folder or a zip. PDFSifter lists the PDFs it finds before anything uploads, so you tick only the ones you want. Progress updates live, and you can cancel at any time without losing finished results.
If a few documents fail, retry just those in the same job and pay only for their pages.
Step 5
Review what’s uncertain, then export
Values below your review threshold (80 by default) are flagged. Step through them with the keyboard, fix what’s wrong, and mark a made-up value “Not in this paper”.
Export to CSV, Excel, JSON or a PDF report. Reviewed documents can go straight back into training.
What it reads
The PDFs you actually have
Digital PDFs
Papers, reports and exports with a text layer, including their tables.
Scans and photos
Pages without text are read with OCR automatically.
Filled-in forms
Answers typed into PDF form fields, ticked checkboxes and chosen options.
One row or many
A paper can give one record, or a chart can give one row per visit.
Password-protected PDFs need to be unlocked before upload.
Try it on your own papers
Your first 200 pages every month are free. Setting up a template takes about 15 minutes.