How AI data extraction works
The tool first turns the document into text: the text layer of a PDF is read directly, and pages that are scanned images, as well as JPG, PNG and WebP files, go through OCR. The text is then given to a language model together with a JSON schema that lists the fields to fill, and the model is constrained to answer with JSON that fits that schema.
Values are taken from the document only. When a field is not stated, it comes back as null or an empty list rather than a guess, and dates and amounts are normalised so the result can go straight into a spreadsheet or a database. The model runs on FileConvertLab's own servers; see the privacy policy for how uploaded files are handled.
Ready-made templates
Five templates cover the documents people most often retype by hand. The invoice template returns vendor, customer, number, issue and due dates, currency, subtotal, tax, total and every line item. The resume template returns contacts, skills, work history and education; the contract template returns parties, dates, term, payment terms, governing law, obligations and signatories.
The product spec template collects model numbers, dimensions, materials, technical parameters and certifications from datasheets, and the academic paper template returns title, authors, venue, DOI, abstract, keywords, methods and key findings. For tables that should stay tables, PDF to Excel is the better tool.
Your own schema
When none of the templates fits, choose Custom schema and describe the fields in JSON Schema: an object with properties, each with a type such as string, number, boolean, array or object, and an optional description that tells the model what to look for. Nested objects and lists of objects are supported, up to four levels deep and 60 fields.
Every list is capped at a fixed number of items and every text field at a fixed length, so the answer always finishes. Keep field names short and descriptive: the model sees the names and descriptions, so "delivery_date" with the description "YYYY-MM-DD" works better than "d1". The same schema can be sent through the API for automated pipelines.
Limits
All pages are read, up to the page limit of your plan. If a very long document runs out of processing time, the result covers the pages read so far and says so. Files can be up to 50 MB. Scans are recognised with OCR first, so a clean, straight scan gives better results than a skewed photo.
Language models can misread or miss values, especially on low-quality scans and unusual layouts. Check important figures against the document before using them. To make a scan searchable instead of extracting fields, use OCR PDF.