docRefineryAlpha

One pipeline, from raw file to validated field.

docRefinery ingests any document format, extracts against a schema you control, and attaches proof to every value it returns.

Request accessSee how it compares
Page inFields outFirst sight

Illustrative route

What happens to your documents.

The first page from a sender is learned once. Every page after it is remembered, not relearned.

IngestPDFs, scans, photographs and spreadsheets, sent through the API.
SegmentationA long file is separated into the documents inside it, and each is recognised for what it is.
TrainingA layout it has not met before is learned once, properly. Not once per document.
Re-trainingWhen a sender changes their form, it is learned again. Nothing for your team to schedule.
RecallRemembered, not relearned. The next page from that sender is recognised on sight.
ExecuteOnly the fields your organisation approved come back, each with a confidence it had to earn.

Define a schema. Get back typed, cited values.

The extraction call carries no field list at all. What may be extracted is account configuration, so a caller cannot ask for fields their organisation never approved.

Request

No field list in the call. What may be extracted is your account’s configuration.

Response · result.response

invoice_number"INV-2291"· 0.97
supplier"Halden Paper Co."· 0.95
total_due"1,284.00"· 0.93
due_date"2026-11-02"· 0.62
po_numbernull· null

Each item is { text, alias, field_confidence }. Under your threshold goes to a person; not on the page comes back null.

Automate the confident cases. Route the rest.

A field either clears the threshold you set or it goes to a person. Nothing ships silently on a number nobody chose.

0.85

4 ship on their own, 2 go to a person. Nothing ships silently on a number nobody chose.

Illustrative readings for one sample invoice.

  • Invoice number
    0.98
    Ships
  • Supplier
    0.95
    Ships
  • Total due
    0.91
    Ships
  • Tax
    0.87
    Ships
  • Payment terms
    0.79
    To a person
  • Due date
    0.64
    To a person
Control
01Confidence thresholds you setDecide per field, per document type, what counts as safe to auto-ship versus what needs a human look.
02Reviewers see the source, not just the valueComing soonThe original page renders next to the extracted field, with the cited region highlighted.
03Corrections train the next runReviewer edits feed back into evaluation, so accuracy on your document types keeps improving.

Run it where your data already lives.

Private VPC and air-gapped on-premises deployment run today on the self-hosted backends. That is a capability, not a compliance certificate — SOC 2, SSO/SAML and data-residency commitments are a separate question, and one we do not answer yet.

Your networkdocRefinery cloud

Managed cloud

Fully hosted and scaled for you. The right default for most teams.

Pages are read inOur cloud

Your networkYour cloud account

Private VPC

Runs inside your cloud account. Documents never leave your network boundary.

Pages are read inYour cloud account

Air gap · no outside connectionSelf-hosted models

Air-gapped, on premises

On self-hosted models that run entirely inside your network, with no outside connection.

Pages are read inYour own hardware

Recent updates.

Dated entries, newest first.

DateEntry
Handwriting recognition accuracy improved on cursive and mixed-script forms.Jun 2026
Added native connectors for Snowflake and Google Drive.May 2026
Schema inference now supports nested and repeating table structures.Apr 2026
Review queue added bulk approval for fields above a confidence floor.Mar 2026

Bring your first document.

Get an API key, send a real file, and see grounded output in minutes.

Book a demoHow the pipeline works