MIT licensed · groundtruth-verify

Ground truth for every document

Turn any document — scans, PDFs, tables, entire folders — into structured data you can trust. No templates at ingest: every file is read completely first, what you use it for comes after.

structured.json
invoice_noINV-482199%
date2026-07-0998%
vendorAcme Corp97%
total$12,480.0099%
tax_id84-291033796%

Private by design

Your documents run on a dedicated instance. Nothing is pooled, shared, or used to train models. Ever.

No templates at ingest

Every file is read completely and faithfully first — text, tables, visuals, the lot. What you use it for is decided after, never baked into extraction.

Every value, cross-checked

Independent reads are diffed cell-by-cell and verified back against the source image. Disagreements are named and flagged — never silently guessed.

No per-page fees

Flat, predictable pricing. Process a million pages without a metered bill scaling against you.

Open source

The verification layer is open

Cross-read table diffing, VLM bounding-box grounding, and NLI-based hallucination rescue — three standalone pieces pulled out of Groundtruth's pipeline, MIT licensed. Use them in your own extraction or RAG system, no strings attached.

github.com/gauravmeena41/groundtruth-verify →
pip install groundtruth-verify
from groundtruth_verify import verify_table_region

read_a = "| Acct | Amount |\n|---|---|\n| LN-1 | 1,00,000 |"
read_b = "| Acct | Amount |\n|---|---|\n| LN-1 | 100000  |"

result = verify_table_region(read_a, read_b)
# {"has_table": True, "diff_count": 0, "cell_diffs": []}
# "1,00,000" vs "100000" normalize equal — not a real diff.

From document to data in three steps

One pipeline for clean digital PDFs and the worst photocopies — routed automatically, verified at every step.

01

Split & read

A bundle of files is broken into its logical documents and labeled, then every page is read — text, tables, and visuals like signatures, stamps, and maps.

02

Verify & link

Each read is checked back against the source image, then entities are matched across every document in the case — with conflicts surfaced, not hidden.

03

Review in minutes

A guided review — scan beside data, click-to-source, flag-jumping — brings each case down to 5–10 minutes, not three hours.

What comes out the other side

Not a mockup — the actual verification queue and entity model Groundtruth produces on real documents.

Cases reviewed in 5–10 minutes, down from three hours
Human Review Queue — why flagged

Every flag names its exact reason — nothing is silently guessed.

cross_read_table_disagreement

1 table cell(s) differ between reads — review

cross_read_disagreement

OCR vs ensemble disagree (agreement 0.07) — candidates surfaced for review

ocr_confidence

confidence 0.40 below floor 0.6

ocr_gibberish

repeated-character run

ocr_blank_page

OCR produced no text for this page

Canonical model — auto-typed entities

Every document is parsed into typed, structured entities automatically.

PERSON
Jane A. WhitfieldR. K. SharmaPriya Nair
ORG
Meridian Finance CorpNorthgate Legal LLP
MONEY
$12,480.00₹7,30,000$860.00
IDENTIFIER
INV-4821REG-88231LN-2024-0917

Built for the teams drowning in paper

If your documents are sensitive, high-volume, or messy, Groundtruth is built for you.

Legal

Sanction letters, KYC, deeds — flag owner-name mismatches against KYC, missing title deeds, deceased-owner records, and export claim amounts.

Finance

Vendor invoices — 3-way match purchase order, invoice, and receipt, then export straight to accounting.

Records

Archives and back-files — digitize decades of paper into searchable, structured data.

Healthcare

Intake forms and lab results — capture values and units, flag abnormal readings, without sending PHI to the cloud.

Security & privacy

Your documents never leave your control

Cloud OCR APIs pool your most sensitive documents on shared infrastructure. Groundtruth doesn't. Every deployment is isolated, and your data is never used to train anything.

Dedicated instance

Your data is isolated to your own environment — no shared multi-tenant pool.

No training on your data

Documents are processed and returned. They never enter a training set.

Compliance-friendly

Built for teams that can't send sensitive documents to a public cloud API.