Beyond stateless extraction.
Every other extractor reads your five hundredth invoice as though it were the first. docRefinery learns a vendor’s layout, keeps it, and reads every later page with what it already knows.
Drop a document into the lens
Or click a sample. Then drop the same one again to watch it remembered.
Structured result
Appears here after a drop
Sample documents are synthetic. Nothing you drop is uploaded.
One engine, every step of the paperwork.
Extract, classify, split, govern and review, in one pipeline that gets quicker with every sender it meets.
Every value, with the evidence it came from.
Fields come back structured and grouped, each with a confidence and its place on the page. Line items come back as rows.
- Line items as rows
- Confidence on every value
- 70+ languages and scripts
- Invoice number
- NWF-204180.998
- Issue date
- 2026-09-140.995
- Bill to
- Aster Components GmbH0.988
- Total due
- 12,476.00 EUR0.997
- PO reference
- PO-77?4120.61
Built to remember.
Most extractors start from nothing on every page. docRefinery studies a sender’s layout once, keeps it, and starts every later page from what it already knows.
Read the signature
Every layout carries one. It is recognised the moment a page arrives.
Remember the layout
The first page from a sender is studied once, properly, and kept.
Return governed fields
Only what your organisation approved, each value with its evidence.
98.02%
of fields extracted correctly in production. The rest were flagged by the confidence score, not shipped silently. Measured in production, not a benchmark.
- 4,000+
- documents a day, sustained
- 20
- types in the governed catalogue
- 785
- field definitions to configure
- 70+
- languages and scripts
Where teams put it to work.
The advantage compounds wherever the same senders keep sending the same forms. Pick an industry to see its paperwork being read: the documents, the fields teams ask for, and why those layouts repeat.
A fixed supplier list sending the same forms every month is the best case this product has.
Documents
Invoices · Purchase orders · Remittance advice · Credit notes
Why the layouts repeat
Your vendor list barely changes. A few hundred layouts cover years of invoices, and each one is learned the first time it arrives.
Your document types, trained for you.
When the catalogue does not cover yours, we train a classification model on your samples, host it, and keep it current. At no charge.
Generated documents · illustrative
Invoice
14 sorted
Bank statement
9 sorted
Bill of lading
7 sorted
Unrecognised
5 for a person
A type the catalogue does not know.
Freight settlements keep arriving and nothing recognises them, so they pile up for a person to sort by hand.
Try a document
Drop one onto the line and watch it decide.
- Training a classifier for your types
- No charge
- Hosting and running it
- No charge
- Retraining as your documents change
- No charge
- Help with fields, thresholds and rollout
- No charge
- Extraction volume once it is live
- Your normal plan
Everything around the engine.
The details that decide whether extraction can be trusted in production.
Present, ambiguous or absent. Never guessed.
It answers only what the page will vouch for. Whatever it cannot stand behind is held back, or quietly handed to a person.
70+ languages and scripts
Left to right, right to left, and the alphabets in between.
PDFs, scans, photos, spreadsheets
Born-digital or photographed at an angle on a phone.
Can be lowered by review. Never raised above what the page supports.
A confidence you cannot talk up
Every score is earned from the page. Below your threshold, it goes to a person.
curl https://api.docrefinery.com/v1/extract \ -H "Authorization: Bearer $DOCREFINERY_KEY" \ -F "[email protected]" \ -F "document_type=invoice" # that is all. The fields are yours to govern.
An endpoint that takes no field list
What may be extracted is your organisation’s configuration, not an argument on the call. A caller cannot ask for fields that were never approved.
Run it where your documents already live.
Private and air-gapped deployment run today. That is a capability, not a certificate: SOC 2, SSO/SAML and data residency are a separate question, and one we do not answer yet.
Managed cloud
Fully hosted and scaled for you. The right default for most teams.
Pages are read inOur cloud
Private VPC
Runs inside your cloud account. Documents never leave your network boundary.
Pages are read inYour cloud account
Air-gapped, on premises
On self-hosted models that run entirely inside your network, with no outside connection.
Pages are read inYour own hardware
Against the alternatives, including where we lose.
Every competitor here sells a stateless parse: send a document, get structure back, and start from nothing the next time the same layout arrives. This table measures that, and the rows where they are ahead of us.
Memory and governance
Trust
Coverage and deployment
Where we lose, who is closest, and how this was sourced
Where we lose. Five rows are ours to lose, and visual citations are the sharpest: Reducto, Unstract, DocuPipe and Textract draw the box on the page today, and we do not. DocuPipe is strong where trust is bought, with per-field confidence, clickable citations, a visual review interface, and SOC 2 Type 2, ISO 27001, GDPR and HIPAA with a BAA. Unstructured.io owns retrieval pipelines we do not have.
The closest in spirit. Extend markets self-improving document AI, ships evaluation, schema versioning and a review interface on every tier, and publishes a 99.2% accuracy figure. It improves a schema against test sets; it does not remember a layout. Where its capability is not separately verified, the cell says so.
Two different companies. Unstructured.io is an open-source ETL and retrieval ingestion platform; Unstract is Zipstack’s open-source LLM extraction platform with Prompt Studio. The names are one letter apart and routinely confused.
Competitor claims are sourced and dated in our landscape review, and checked against each vendor’s own published material. Re-check before quoting: published pricing and capability move.
Questions, answered plainly.
Including the ones where the answer is not yet.
It is recognised on sight. The first page from a sender is studied once; every page after starts from what was already learned, so it is quicker and steadier than the first.
It says so. The field comes back as null rather than a plausible guess, and anything uncertain comes back with a score under your threshold so a person can look.
PDFs, scans and photographs, and spreadsheets. It is not a general converter for presentations or the long tail of office formats, and it is not a retrieval pipeline.
We train a classifier on your samples and host it ourselves, and help you set up the fields and thresholds. Training, hosting, retraining and that help are free; you pay only for your normal extraction volume.
Yes. Private VPC and fully air-gapped on-premises deployment run today, on models hosted inside your network.
Not yet, and we would rather say so than imply it. Self-hosting is a capability; those certifications are a separate question.
Bring one vendor's invoices.
The first document teaches us the layout. Every one after it is recognised on sight.