Frequently asked questions
The short answers. For step-by-step detail see the documentation, and for plans and billing see pricing.
Getting started
What does dataextractor.io actually do?
It turns documents into structured fields you can use. Upload an invoice, purchase order, statement or similar document and it returns named values — invoice number, supplier, dates, totals, line items — rather than a wall of text. Every value is shown against the page it came from, so you can check it before you trust it. Where it differs from plain OCR is the feedback loop: when you correct a value, that correction is used to rewrite the extraction instructions for that field, so the same mistake stops happening on future documents of the same type.
How do I run my first extraction?
Open the Extractor from the top nav and upload a document — or pick one of the demo documents, which never cost a credit. The wizard walks you through four steps: it confirms what the document is, proposes the fields worth pulling out, runs the extraction, and shows you the results next to the document for review. You do not need to configure anything up front or supply training data.
Getting started guide →Do I need to train it or upload samples first?
No. Schema detection reads your document and proposes the fields itself, so the very first document works with no setup. Training happens implicitly afterwards — each correction you make teaches the system, and accuracy climbs over the first two or three documents of a new type.
How long does an extraction take?
A standard invoice typically finishes in about 30 to 60 seconds. Longer documents and documents with large line-item tables take proportionally more. Batch runs process documents in parallel, so a folder is much faster than running each file by hand.
Documents & formats
What file types and sizes are supported?
PDF — both born-digital and scanned — plus JPEG and PNG, up to 25MB per file. Multi-page documents are fine: all pages are processed together as a single extraction, and a multi-page document still counts as one document.
Does it work on scanned or handwritten documents?
Scanned documents work, but quality matters — scan at 200 DPI or better before spending time tuning, because OCR noise on a poor scan will cap accuracy no matter how good the instructions are. Handwriting is more mixed: printed identifiers such as ID numbers, dates and amounts extract dependably, while cursive names often do not, and re-running or switching models will not reliably fix a hard hand. On those documents it is better to anchor on the printed identifiers and confirm names by eye.
Does it handle non-English or mixed-language documents?
Yes, including documents that mix scripts within a page. If the document uses local terms you want mapped to your own labels, tell the copilot the mapping, save it as ground truth and run 'improve accuracy' — that bakes the mapping into the field's instructions so it applies to every future document of the type, instead of only the one in front of you.
What if one upload contains several documents?
The system detects logically separate documents inside a single upload — for example a bundle of supplier quotes attached to one request — and returns each one separately in the results. If a bundle is being treated as a single document when it should be split, say so in the copilot and it can help correct the split.
Accuracy & the learning loop
How accurate is it?
Expect to correct three to five fields on the first document of a type you have never run before. After one round of corrections and one learning round, accuracy typically passes 90 percent on the same layout; after two rounds most teams reach 95 to 99 percent on consistent layouts from the same supplier. Born-digital PDFs do best; scan quality is the main variable on everything else.
Accuracy dashboard →A field came out wrong. What do I do?
Correct it in place on the review step — click the value and type the right one — then save it as ground truth. That fixes the document in front of you. To stop it recurring, run 'improve accuracy' afterwards: the learning loop compares your corrections against what was extracted, works out why each field failed, and rewrites that field's extraction instructions. The next document of the same type gets the improved instructions.
Troubleshooting guide →What is the learning loop and does it cost anything?
It is the step that converts your corrections into better extraction instructions. It compares the saved ground truth against the extraction output, identifies what kind of error each field had, rewrites that field's instructions to address it, and re-checks the document. A learning run costs one credit, the same as an extraction. The instructions it produces are visible in the Fields to Extract panel and you can edit them by hand — it is deliberately not a black box.
Two fields keep returning the same value. How do I separate them?
That is a schema conflict — usually two fields whose descriptions point at overlapping parts of the page, like invoice date and due date, or subtotal and total. The fix is to make each description explicitly exclude the other: 'the invoice issue date, not the due date or the service period start'. The copilot can write both descriptions for you if you describe the confusion.
How do I know where a value came from?
Every extracted value is highlighted on the document itself with a bounding box, so you can see the exact region it was read from rather than taking the value on faith. Reviewing against the highlight is the fastest way to catch a value that looks plausible but was read from the wrong place.
Workflows, batch & matching
Do I have to repeat the setup for every document?
No — that is what workflows are for. When you finish a document in the Extractor, save the configuration as a workflow. Every later document of that type reuses the saved schema and the tuned instructions, so it runs without any schema review. The workflow is also what the API and batch runs point at.
How do I process a lot of documents at once?
Use Batch. Upload the whole set, confirm the schema once, and the documents run together; results come back as one combined export. Tune the document type in the Extractor first — batch is for volume once the type already works, not for figuring out a new layout.
Can it match extracted lines against my own product or material list?
Yes. Upload your catalogue, material master or price list as Excel or CSV on the Reference Data page, then run Matching against an extracted dataset. That turns free-text line descriptions on a document into real records in your own system, which is usually the step that makes the extracted data actionable rather than merely correct.
Can a team review extractions before the data is used?
Yes. Batch and workflow runs seed a Review queue where a person confirms or corrects each document before it is trusted downstream. Reviewer access or higher is required, and corrections made there feed the same learning loop — so a team reviewing together improves the system faster than one person working alone.
API & integrations
How do I get an API key?
Create one on the Developer page, reached from More → API in the top nav. Keys start with dex_, can be named so you can tell production from staging, and can be revoked individually without disturbing your other keys. Send the key in the X-API-Key header — it is not a Bearer token.
API reference →Can I call it from my own code?
Yes. POST a file to the extraction endpoint with your workflow's slug and you get the structured fields back in the same response, or submit it as a job and poll if the document is large. The response includes each field's dot-path, its value and a confidence score, plus a version grouped into headers and line-item arrays. The API is served at api.dataextractor.io.
API reference →Does it connect to my ERP or accounting system?
Extraction results can be exported as CSV, Excel or JSON, or pushed to a webhook when a document finishes, which covers most systems that can accept a file or an HTTP call. For a direct connection, the REST API returns structured JSON you can write straight into your own system. See the integrations guide for the specifics.
Integrations guide →What formats can I export to?
CSV, Excel and JSON, or a webhook that receives the result as it completes. Exports are available from the Extractor's final step for a single document and from the Results step of a batch run for a whole set.
Security, data & accounts
What happens to my documents?
Documents are stored so you can revisit a past extraction from the Datasets page and so the review and learning steps have something to compare against. Access is scoped to your account, and admins and owners can see security-relevant activity on the Audit Log page. The privacy policy and the DPA cover the specifics, including retention and deletion.
Privacy policy →Can I control what team members can do?
Yes — access is role-based. The Review queue and Reference Data need reviewer access or higher, and the Audit Log is restricted to admins and owners. Pages you do not have access to are hidden from the navigation, and the restriction is enforced on the server as well, not just in the interface.
What can the in-app assistant do?
It works alongside the Extractor rather than just answering questions: it can add, rename or remove schema fields, correct extracted values, start a re-extraction, run the learning loop, and explain what a step does or where a feature lives. Ask it in plain language — 'remove the fax field', 'the PO number should be PO-2568-00847', 'where do I get an API key' — and it acts on the interface directly.
Still stuck?
The assistant inside the Extractor answers product questions as you work — and can edit fields, fix values and run the learning loop for you. Or email us.