ProductDataForAI — Product Intelligence Hub

01Data first, AI second

The facts are settled before any model sees them.

Many AI applications read documents in full text and leave it to the language model to find the facts. The hub reverses the order: the facts are checked and structured beforehand. The AI receives them and only phrases the answer.

Full-text AI app Product Intelligence Hub Basis PDFs and texts, cut into chunks Attributes with value, unit, language and origin “Which valves take more than 10 bar?” The model estimates from the chunks it happens to see Counted exactly over typed values, without a model Similar products Name variants are easily mixed up Every product is its own record with a unique key Same question, twice The answer may differ Same data, same result Traceable Points to a passage of text Every value names its source system A change in the source Re-read the documents and hope the old version is gone Changed in one place, valid everywhere Role of the AI Searches, selects and phrases Phrases, from facts it was given

01 · deterministicConnect

Every source attached where it lives.

02 · deterministicCleanse and check

Notation, units, duplicates, gaps.

03 · deterministicLink and enrich

Inheritance, references, facets, markets.

04 · AIAnswer

The model phrases the answer from checked facts, with the source.

02Clean data

Errors in product data rarely show.

A filter quietly returns fewer hits, a catalog drops a unit, an assistant names the wrong product. None of it raises an alarm. This is why the hub checks the data before it is passed on.

AI does not clean up bad data. It makes it convincingly wrong.

Invisible charactersA non-breaking space inside a value, and the filter no longer finds the product.
Ambiguous valuesThe same text means different things in different attributes. Without structure, an AI guesses which is meant.
Duplicate namesTwo product series share a name but serve different markets. Only the links tell them apart.
Missing formatsWithout a number format on the attribute, unit and operator disappear, and the value still looks right.
Truncated listsA comma inside a value, and a list quietly ends after its first entry.

03Mixing by linking

No system knows the whole product.

In the hub the parts are linked, not copied. Every value keeps its origin, and the hub adds what no source provides. The same article as in the harmonization table, now complete.

from a source system added by the hub

Ball valve, stainless steelART-4471-A

Commercial Article numberART-4471-AERP List price48,90 EURERP Description Product textFull-bore ball valve for drinking water and industrial mediaPIM Languagesde · en · frPIM Technical ClassBall valve (ETIM)Classification Thread · max. pressureG 1/2″ · 16 bar, inherited from the seriesInheritance Material1.4404, with its properties from the linked materialReference Media Image · data sheetProduct photo, data sheet (PDF)DAM Meaning Filter facetsMedium = water · Connection = threadCalculated MarketsDE · AT · CHCalculated WeightingPressure and material count more for search and AIRule

Example article. Values are illustrative.

04Adding meaning

Maintained once, valid wherever it belongs.

Three kinds of link carry a value to every place it belongs. That is how linked data gains meaning: each value knows its product, its attribute and its context.

Inheritance

Series BV-4400Thread G 1/2″ · 16 bar

passes on to

Article ART-4471-Ashows both values without owning them

What a series shares is kept at the series. Every article inherits it, and one change reaches all of them.

Reference

Article ART-4471-AMaterial → 1.4404

refers to

Material 1.4404resistance, approvals, notations

An article points to another object, even in another source system. Shared knowledge is kept once and used by many articles.

Mirror

Article in the catalogthe original with all values

mirrored into

Article in the web structureall values visible, read only

The same article appears in several structures, such as catalog, web and print. It is maintained in one place.

Data

Single values from the sources, each on its own.

16 · 1.4404 · ART-4471-A

Information

Linked, classified, with unit and language. Every value knows which product and attribute it belongs to.

max. pressure 16 bar · material 1.4404

Knowledge

Facets, markets and weights the hub calculates. With them, finder and AI answer questions no single source can answer.

“Stainless valves for DE above 10 bar?”

05Connecting sources

Every source is a domain.

A domain describes one connected system: how to reach it, what it is for and how it is set up. A new source is a new domain, not a rebuild.

Anatomy of a domain

Connection: Protocol and address, e.g. ERP service layer, PIM REST, file store
Purpose: Storage, index, print, translation, AI, workflow …
Settings: Options per domain, adjustable without new software
Links: Domains use each other: an index reads values from the source, an assistant searches the knowledge base

erp.articlespim.catalogdam.mediaweb.finderprint.indesign

Existing systems stay in charge. The hub reads them where they are.
Every value can name the domain it came from.
Domains can be switched on and off, or limited to particular machines.
Write access is a setting per domain: what an assistant may only read stays read-only.

06Automation

Built-in scripting connects the domains.

A script step can call any domain, check its result and hand it to the next. New workflows are configuration, not programming.

Example · on request A question finds the right source of knowledge

The system AI classifies the question: installation or product data.
A rule picks the matching domain.
The question goes there; the answer comes back with its sources.

Example · scheduled Overnight, the knowledge base catches up

A task starts at a set time.
It reads the changed products with all linked values.
The knowledge base is updated; assistants answer from the new state the next morning.

Import and reconciliation Values calculated on read Scheduled tasks Web and print production AI workflows Setting up new customers