Tolva Extract · model DX-2

Documents in, typed JSON out. Measured, not claimed.

Send a PDF, a scan or a phone photo; get back JSON that validates against your schema. 99.4 % of fields come back right the first time, and every field carries its own confidence, so you know which of the rest to check.

invoice.jsonconf
{ "vendor": "Harrow & Pike Ltd.",0.997 "invoice_number": "INV-20417",0.998 "issue_date": "2026-10-03",0.996 "due_date": "2026-11-02",0.994 "bill_to": "Ostrander Studio",0.995 "line_items": [ { "item": "Archival paper, A3",0.991 "qty": 4, "amount": 154.00 }, { "item": "Letterpress plates",0.988 "qty": 2, "amount": 192.00 }, { "item": "Bookbinding thread",0.984 "qty": 10, "amount": 74.00 } ], "subtotal": 420.00,0.999 "tax": 84.00,0.997 "total": 504.00,0.999 "currency": "GBP",0.993 "payment_terms": "net_30"0.989}

Valid · 13 of 13 · 3 below 0.990

Calibration · DX-2 release 2.6

That was one document. Here is the same model on 12 480.

Each release is run against 12 480 held-out documents before it ships, and scored field by field against a stated tolerance. This is the run behind the API today.

FieldAccuracy by batchResult
Invoice totals99.6 % ± 0.2, within tolerance
Dates99.4 % ± 0.3, within tolerance
Line items98.7 % ± 0.4, within tolerance
Addresses98.9 % ± 0.4, within tolerance
Tax IDs99.8 % ± 0.1, within tolerance
Currency codes99.7 % ± 0.2, within tolerance

Stamped: pass, within tolerance. Signed by Tolva Metrology. Certificate TX-2611-0412, Issued 16 Nov 2026.

Every value comes back with its evidence.

0 to 1

A confidence on every field

Set a threshold. Anything under it goes to a person, with the field marked on the page where it was read.

Your schema

Typed, or an error that says why

Send a JSON Schema; get values that validate against it, or an error naming the field that failed.

TX-2611-0412

A certificate on every response

Each response names the calibration certificate of the release that produced it, so any value can be traced to how that model was measured.

Characteristics · DX-2 r2.6

What it reads well, and where it doesn’t.

Typical values from the held-out set for this release: 12 480 pages it never trained on. Conditions sit next to every figure, so you can hold us to them.

Fig. 1 · n = 12 480 pagesField accuracy against scan resolution
92949698100100200300400500600dpi% fields correct99.4 % at 300 dpi
Fig. 2 · n = 1.6 M fieldsStated confidence against how often it was right
0.50.60.70.80.91.00.50.60.70.80.91.0statedobservedsays 0.90, right 90 % of the time

Table 1 · % of fields correct · 300 dpi · r2.6

Field accuracy by document type

Document type

Pages

Min

Typ

Max

Invoices

4 210

99.1

99.6

99.9

Receipts

3 060

98.4

99.2

99.7

Bank statements

2 140

99.4

99.8

99.9

Purchase orders

1 820

98.9

99.5

99.8

Handwritten forms

1 250

95.6

97.9

99.0

All documents

12 480

95.6

99.4

99.9

API · one endpoint

A file and a schema in. Typed JSON out.

One request per document, answered in the same call up to 20 pages. Every field under your threshold comes back in review, with the page and the box it was read from.

Fig. 3 · Pin-out · POST /v1/extract

Four in, four out

file

PDF or image

1

schema

JSON Schema

2

threshold

0 to 1

3

locale

BCP 47 · optional

4

DX-2

POST

/v1/extract

r2.6

5

values

typed per schema

6

confidence

0 to 1 per field

7

review

under threshold

8

certificate

TX-2611-0412

Request

curl

curl https://api.tolva.example/v1/extract \
  -H "Authorization: Bearer $TOLVA_KEY" \
  -F file=@invoice-0412.pdf \
  -F schema=@invoice.schema.json \
  -F threshold=0.9

Response

200 · 1.8 s

{
  "values": {
    "total": 1284.50,
    "due": "2026-12-01",
    "vat_id": "DE 811 907 980"
  },
  "confidence": {
    "total": 0.998,
    "due": 0.973,
    "vat_id": 0.861
  },
  "review": ["vat_id"],
  "certificate": "TX-2611-0412"
}

Ordering information

Priced per page read, with the limits printed.

Every plan reads with the same model and comes back with the same certificate. What changes is volume, file size, rate and how long we keep your files.

Table 2 · Ordering information · prices in USD, before tax

TLV-FREE

Free

$0

for the first 500 pages each month

Pages

500 / month

File size

20 MB

Rate

2 req/s

Retention

7 days

Support

Community

TLV-USAGE

Most teams

Usage

$12

per 1 000 pages, billed monthly

Pages

No cap

File size

50 MB

Rate

20 req/s

Retention

30 days

Support

E-mail, 1 business day

TLV-VOLUME

Volume

Custom

from 1 M pages a month

Pages

From 1 M / month

File size

100 MB

Rate

Dedicated

Retention

Set per contract

Support

Named engineer, 99.9 % SLA

Pages are counted once per document page read, whether or not a field lands in review. Unused Free pages don’t roll over.

Revision history

Each release ships with its certificate.

A new model version is measured on the same held-out set before it reads a single customer page. If the number goes down, it doesn’t ship.

Table 3 · Revision block · accuracy on the held-out set

Rev

Date

What changed

Certificate

Acc. %

r2.6

Handwritten forms up from 96.2 % to 97.9 %. Confidence recalibrated on 1.6 M fields.

TX-2611-0412

99.4

r2.5

Bank statements with merged cells read as tables. Receipts under 150 dpi go to review.

TX-2609-0388

99.2

r2.4

HEIC files accepted. Thresholds can be set per field as well as per request.

TX-2607-0301

98.9

r2.3

Evidence boxes returned for every field, not only the ones in review.

TX-2606-0217

98.6

Start

Send it your worst page.

The first 500 pages each month are free, with no card. Run your hardest documents through it and read the certificate that comes back.