New: Extraction Engine

Fields, not free text. From texts and photos.

You say which fields you need. The Extraction Engine reads them from text and up to four photos and returns each field typed: with a confidence and the probability that there is no value at all.

Extraction · sovr-extract-1.0
Photo: back of a package

Fields

  • Best before · date2027-12-12
    Confidence 0.98
  • Lot code · textL5123A0457
    Confidence 0.93
  • Net weight · number500 g
    Confidence 0.97
  • Seal intact? · booleanyes
    Confidence 0.91

Illustration · each field is its own pass · read in Germany

At a glance

Typed fields

Eleven field types from text to estimate. Your software gets values, not prose.

Confidence per field

Every field says how sure the value is and how likely there is none.

Location in the photo

On request, a box shows where the value sits in the image.

Operated in Germany

At STACKIT. Nothing is stored, nothing is trained.

Eleven field types

  • textCodes, names, numbers – exactly as printed
  • numberAmounts, quantities, sizes – with a unit
  • dateBest-before or invoice date – as an ISO date
  • timeTimes of day
  • datetimePoints in time with date and time
  • durationDeadlines and terms
  • listSeveral entries: line items, ingredients, numbers
  • booleanYes or no, with a probability
  • enumOne of up to 25 fixed options
  • boxWhere something is in the photo
  • estimateA value within a range, with median and spread

We deliberately do not offer forecasts over several time steps.

What each field returns

value
The value read, in the field's type – or null if there is none.
confidence
How sure the value is, between 0 and 1.
p_null
How likely the field is empty or unreadable.
guess
The best value even when value is null – for a human to check.
span · box
Where the value is: a character range in the text or a box in the photo.
warnings
Notes, for example when a value equals the example in your instructions.
"lot_code": {
  "type": "text",
  "value": "L5123A0457",
  "guess": "L5123A0457",
  "p_null": 0.02,
  "confidence": 0.93,
  "box": { "image": 0, "image_name": "rueckseite",
           "x1": 412, "y1": 988, "x2": 731, "y2": 1032, … },
  "warnings": [],
  "calibration": { "status": "uncalibrated" }
}

What for

Anti-counterfeiting for packaging

Read the lot code and best-before date from a photo of the back and match them against your data.

Invoices and delivery notes

Number, date, amounts and line items as fields, ready for your accounting.

Forms

Ticked or not, date, signature present – field by field.

Identifiers in scans

File, customer or serial numbers from scanned pages.

Sovereign operation

  • Fields are read in Germany, at STACKIT. According to STACKIT, its model serving falls within the scope of its BSI C5 attestation.
  • We do not store text, photos or fields. Billing is by tokens.
  • No training on your data.
  • The processing location is in every answer.

Fixed versions

You call sovr-extract-1.0, and exactly this version answers: runtime, prompt and scoring are pinned. Each version stays available for at least 90 days, this one until at least February 4, 2027.

The same request gives the same value; probabilities can vary in the third or fourth decimal place. The model weights are held by the operator.

Our measurement (warm, one photo at 1600 pixels, four fields at once):

854 ms at the median · 1,441 ms in 95 % of runs

Each field is its own pass. These are measurements, not a guarantee.

What you should know

Not yet calibrated

Confidence and p_null are not calibrated yet. Set your thresholds on a test set of your own.

Measured on real products

Twelve packaging photos: date right in 12 of 12, lot code in 9 to 10 of 12. Codes read wrongly had a confidence of at most 0.76.

Small print needs resolution

For printed codes, about 1600 pixels on the long edge. The whole request may be at most 4.4 MB.

No example values in instructions

When the print is unreadable, models like to copy an example from the instructions. The answer then warns with matches_instruction_example.

Operator limit

Today 45 requests per minute for this version, and each field counts as one. If you need more throughput, talk to us.

For developers

One endpoint, JSON in, JSON out. Photos travel as Base64 in the request body.

curl https://sovrgpt.com/api/v1/extractions \
  -H "Authorization: Bearer $SOVR_KEY" -H "content-type: application/json" \
  -d '{
    "model": "sovr-extract-1.0",
    "images": [{ "name": "rueckseite", "data": "<base64>" }],
    "fields": {
      "best_before": { "type": "date", "instructions": "Best-before date" },
      "lot_code": { "type": "text", "instructions": "Lot code printed next to the date", "locate": true }
    }
  }'

Decision Engine: decisions via API

You send a text or state plus fixed options. You get back a probability for each option and a recommendation your software can act on directly.

Go to the Decision Engine

Prices on request. Confidence and p_null are not calibrated yet. We keep measuring and will let you know when limits or versions change.

Show us your use case.

Request access
Extraction Engine – fields from texts and photos via API | SovrGPT