Every field, typed. Every field, scored.

An OCR hands you a string and wishes you luck. Gemina returns headers, line items and your own schemas as JSON your code can trust — with confidence on every value.

See the full response
A business document resolving into structured JSON — vendor, totals, currency, and line items as typed fields
Confidence

It tells you what it isn't sure about.

A single model call returns a confident answer whether or not it is right. Gemina scores each field so your review queue routes itself.

Three extracted fields scored with high, medium and low confidence, with the missing low-confidence Tax ID routed into a one-field review queue

Per field, not per document

High, medium or low on every value — so one doubtful line item doesn't discredit the whole extraction.

It refuses to guess

A field that isn't there comes back empty and flagged, not filled in with something plausible.

Corrections feed back

Submit verified values for an extraction and get a per-field comparison of what was right, wrong or missed.

Human verification

When accuracy matters, put a person in the loop.

Drop a complete review step into your product. Your user sees the source document beside every extracted value, corrects only what needs attention, and sends the audited result straight back to Gemina.

Gemina verification interface with an invoice beside editable extracted fields, a missing value marked for review, an edited total and a line-item table
Headers, line items and custom template fields all use the same review flow.
  1. 01
    Route attention

    Hide high-confidence fields so reviewers focus on uncertainty.

  2. 02
    Correct in place

    Edit typed fields, add missing values or fix whole table rows.

  3. 03
    Submit once

    The verified payload is stored with the original extraction.

  4. 04
    Keep your schema

    verifiedValues has the same shape as values.

Document viewer, field controls, table editing, validation, confidence filters and submission logic — already built.

See the drop-in component
Templates

Your schema, built from one sample.

Upload one example, mark the fields you want, and every document of that type afterwards returns the same shape. No labelled training set, no model to fine-tune, no waiting.

Pass the template ID on the call and you get that shape back. Variations of the same document type still land in it, so a vendor redesigning their layout doesn't break your integration.

The schema is yours — editable, exportable, and searchable and filterable alongside everything else you extract.

An onboarding-form template builder with selected full name, customer ID and start date values mapped to reusable typed fields
Depth

Line items are where OCRs give up.

Ten years of invoices taught us the hard part isn't the total. It's the twenty rows above it.

The extraction details view: a photographed invoice beside its 13 extracted line items, with item code, quantity and unit of measure preserved for every row

Every column above is a field you can filter and search on — not a cell in a picture of a table.

Models

Trade accuracy for speed, per call.

One parameter. No separate endpoints, no separate contracts.

InvictusHighest accuracy
Complex or multi-language documents. Costs the most credits.
PraetorianBalanced
The recommended default for most work.
VeloxFastest
High volumes of straightforward documents. Header extraction starts from 4–6 seconds.

Optional passes — thinking, evaluation, correction and coordinate output — trade more time for more accuracy on hard documents.

See it on your own document.

Full API access, all models included. No credit card required.

Try the Live DemoStart Free Trial