Franz Bettag’s own engineering project combines a vision model with strict data validation and a DATANORM exporter. The public demo makes each step from the supplier document to the selected export records inspectable.
EUR 8.50 per 100 pieces is not EUR 8.50 per piece. The reference preserves the price basis. A price on request stays empty.
| Article | Price / basis | Export |
|---|---|---|
MW-101Montagewinkel 40 x 40 mm |
2.40 EURper 1 pcs | Include |
MW-205Schutzkappe, schwarz |
8.50 EURper 100 pcs | Include |
MW-310Dichtband 20 mm |
1.85 EURper 1 m | Include |
MW-420Sonderhalter nach Zeichnung |
On request | Excluded |
DATANORM: 3 articles included. 1 position without a price excluded. All prices exclude VAT.
{
"positions": [
{
"article_number": "MW-101",
"bbox": null,
"currency": "EUR",
"description": "Montagewinkel 40 x 40 mm",
"evidence": "MW-101 Montagewinkel 40 x 40 mm 2.40",
"minimum_quantity": null,
"pack_quantity": null,
"price": "2.40",
"price_quantity": "1",
"price_type": "net",
"uncertainties": [],
"unit": "Stück",
"variant": null
},
{
"article_number": "MW-205",
"bbox": null,
"currency": "EUR",
"description": "Schutzkappe, schwarz",
"evidence": "MW-205 Schutzkappe, schwarz 8.50",
"minimum_quantity": null,
"pack_quantity": null,
"price": "8.50",
"price_quantity": "100",
"price_type": "net",
"uncertainties": [],
"unit": "Stück",
"variant": null
},
{
"article_number": "MW-310",
"bbox": null,
"currency": "EUR",
"description": "Dichtband 20 mm",
"evidence": "MW-310 Dichtband 20 mm 1.85",
"minimum_quantity": null,
"pack_quantity": null,
"price": "1.85",
"price_quantity": "1",
"price_type": "net",
"uncertainties": [],
"unit": "m",
"variant": null
},
{
"article_number": "MW-420",
"bbox": null,
"currency": "EUR",
"description": "Sonderhalter nach Zeichnung",
"evidence": "MW-420 Sonderhalter nach Zeichnung auf Anfrage",
"minimum_quantity": null,
"pack_quantity": null,
"price": null,
"price_quantity": null,
"price_type": "net",
"uncertainties": [
"Preis auf Anfrage; nicht für den Preisexport ausgewählt."
],
"unit": "Stück",
"variant": null
}
],
"reference_kind": "manually_transcribed_reference",
"reference_note": "Manuell gegen das fiktive Original geprüfte Referenz, keine gemessene KI-Ausgabe. Manually checked reference from a fictional document, not a measured AI result.",
"schema_version": "1.0",
"source_kind": "synthetic_demo",
"source_page": 1,
"supplier": "Musterwerk Nord",
"valid_from": "2026-09-01",
"warnings": []
}V 070926Preislisten-Konvertierung Auszug - Werte gegen Quelle pruefen Artikelstamm 04EUR
A;N;MW-101;00;Montagewinkel 40 x 40 mm;;1;0;Stk;240;;;;
B;N;MW-101;;;;0;0;0;;;;;;;;
A;N;MW-205;00;Schutzkappe, schwarz;;1;2;Stk;850;;;;
B;N;MW-205;;;;0;0;0;;;;;;;;
A;N;MW-310;00;Dichtband 20 mm;;1;0;m;185;;;;
B;N;MW-310;;;;0;0;0;;;;;;;;
DATANORM 4 · CP850 · CRLF · EUR
Reference output manually checked against the fictional source. It explains the data flow; it does not measure model extraction quality.
Price lists mix prices per piece, per 100 pieces and per metre. Variants, packaging and prices on request require more than copying a table. A wrongly assigned price can still be structurally valid in the receiving system.
Design and implementation of the document workflow, model integration, output schema, review interface and DATANORM export. This is an independently developed demonstration using fictional documents.
PDF pages are rendered in sequence; images are orientation-corrected and size-limited. Qwen3-VL-8B-Instruct extracts price positions. The inspected demo pins its Qwen revision. Dense grid tables are split into sections while retaining headers and footnotes as context. Original price text is stored alongside the normalized numeric value.
Pydantic validates the JSON schema. Decimal values, quantities and source coordinates are checked. Missing prices, currencies and unclear tax treatment remain visible as review notes. Source excerpts and page references support verification; the location markers are also model suggestions.
Rows are reviewed, edited and individually selected for export. The exporter writes DATANORM 4 V, A, B and optional T records in CP850 with CRLF. Article identifiers, EUR currency, prices excluding VAT and price units must satisfy the export rules.
The prompt treats document content as data. Schema and export checks constrain downstream processing, and CSV cells are escaped against formula execution. Content review and a tested import into the customer’s system remain part of the workflow.
The manually checked reference exports three articles. The fourth has a price on request and is excluded. EUR 8.50 per 100 pieces retains that price basis. This is a reproducible export example, not a measurement of AI extraction quality.
A checkpoint is available after each completed page. Processing pauses when the quota is exhausted; the original document and checkpoint allow it to resume later. Saved JSON can be reviewed and converted to DATANORM without another GPU call. The source document, reference JSON, CSV, DATANORM.001 and export report are linked below.
Quality and business impact are measured on customer test pages against checked references: correct fields, missing rows, correction time, runtime and successful imports. The demo does not publish customer performance figures.
PDF, PNG, JPG or WebP up to 100 MB. PDF pages are processed in sequence, with up to 120 positions per model response and 100,000 positions in the combined export. Failed or incomplete pages remain marked as unfinished. The DATANORM 4 export supports new EUR article records with one general price per article. Tiers, uncertain mappings and missing prices require correction or exclusion.
Live extraction runs on Hugging Face ZeroGPU and depends on the visitor’s quota. The fixed reference remains available independently. Data handling, hosting, supplier formats and target-system integration are specified for each customer deployment.
The fictional source and manually checked reference are hosted on bett.ag. The DATANORM export report records included and excluded rows. The interactive demo exposes the full workflow.
A small vision model is a concrete candidate for document tasks. Tests on your own formats determine the right model size and deployment approach.
