01 · Arrives
A file lands in the format the supplier had, not the one you need.
128 pages of Italian price list, a scan with no text layer, and a spreadsheet where the same bearing is written four ways. Nothing is pre-cleaned, because nothing ever is. Three fields in this batch are unusable and one is blocked outright.
Input: listino_2026.pdf, 45,120 rows, 12 formats
02 · Decoder
It reads tables, prose and drawings alike, in any language.
No template and no mapping file. The Decoder reads the layout as it stands and returns structured records, keeping the source page against every field so the provenance survives the conversion.
Out: 128 records · every field ← source page · units detected per column
03 · Classifier
Each product is placed on a node, with the reason it was chosen.
ETIM, eCl@ss or your own taxonomy. The Classifier reads the whole record rather than the title, scores every level of the hierarchy on its own, and sends the borderline ones to a human instead of guessing.
Out: ETIM EC000380 · 0.94 · reason ← 3 attributes · level 4 at 0.71 → review
04 · Matcher
The core call: is this the same product, or only a similar one?
DN25 is one inch nominal bore. PN40 is a 40 bar rating. Both bodies are AISI 316. Three attributes decided this match and the title similarity was not one of them — which is exactly why a cross-catalog link still waits for review.
Out: same product · 0.97 · 3 deciding attributes · datasheet p.2, row 812, ERP SKU 10-4471
05 · Enricher
Gaps are filled only where a source supports them.
Seven of thirty-two fields arrived populated. The Enricher completes eleven more from the documents you already have, cites the document, page and value for each, and leaves a genuine gap empty rather than inventing a value to look complete.
Out: +11 attributes, each cited · material ← datasheet p.3 · 2 conflicts → review
06 · Resolved
One record your ERP will take, and a reason your team can check.
Four supplier spellings became one master record with a confidence score, the evidence behind it, and a clear statement of what needed review and what did not. The Monitor keeps watching: a changed price or spec is flagged before it reaches a quote.
Out: 6204-Z → DIN 625-1 · 98.4% · sourced · review not required