The product

KaraK, the AI agent platform for your product data

KaraK sources, enriches, matches and validates product data at scale. Every delivered value is linked to its source and supervised. KaraK does not replace your PIM: it feeds it — through Akeneo, an ERP, a dedicated interface, an API or exports.

The name echoes the French shorthand “les caracs”: product characteristics are the common building block across all four links in the chain.

Book a demo Build a product record live

In KaraK's live demo, enter a product and its brand and watch real agents search, cross-check and enrich it — no registration required.

  • in production on a catalogue of 640,000 products
  • +1,000,000 matching operations in one month
  • hosted in France · processed exclusively in Europe

What does KaraK do?

Four links, one chain — each link strengthens the next

Even for matching, KaraK starts with enrichment. A product record equipped with all its characteristics lets agents recognise and prove an exact match, or propose a defensible similar product.

01

Source

Find the raw material where it actually exists.

Research agents start from your product and identify the right sources: manufacturer website, PDF data sheet, catalogue and product photos. Sources are ranked — manufacturer first, then recognised retailers — while questionable sources such as clearance sites and unchecked specifications are excluded.

Crawling is centralised and industrialised, with rate control, error recovery and automatic discovery of new websites by the agents themselves.

web · PDF · photos · every format where product data lives

02

Enrich

Complete every missing attribute according to your schemas.

Every product category has its own schema: expected attributes, units and formats. Extraction agents read the selected sources — text, tables, pictograms and photos — and write each value in the required format: “8.5 litres per minute” becomes 8.5 L/min if that is what your catalogue expects.

Parent/variant grouping avoids paying for the same work twice: values established at model level are propagated to its variants.

tens of thousands of products processed each year

03

Match — exact & similar

Find the same product, or a defensible equivalent, in another catalogue.

Using the enriched record, agents match your products to another catalogue: a competitor, supplier or marketplace. Exact matching finds strictly the same product; similar matching proposes an equivalent with identical or very close characteristics while respecting brand tier — premium with premium, mid-market with mid-market.

This is the kind of matching scripts cannot perform: it requires understanding characteristics, not comparing strings. Use cases include competitive and price intelligence — the higher the match rate, the larger the share of the catalogue whose prices can be managed — as well as substitution, cross-selling and tenders.

+1,000,000 matching operations in one month

See KaraK for price intelligence
04

Validate

Check, harmonise and prove — AI proposes, human oversight validates.

Every written value is checked against its source, harmonised under your rules and linked to the precise place where it was read. Anything that cannot be proven is marked as inferred — never concealed.

Batch processing is supervised end to end, with quotas, automatic error recovery and human control points wherever you need them.

a source or status for every written value

Evidence at value level

How is every value proven?

Every value delivered by KaraK carries two things: a reading explanation — where the data was found and how it was interpreted — and a clickable link to its source: URL, PDF and page, or photo. Anything not read directly is marked as inferred, never concealed.

An inference can come from an indirect reading — recognising the spray head in a photo — or from reasoning between characteristics: if the cartridge is 35 mm and installation uses a single hole, the expected drill diameter is 35 mm. In both cases, the reasoning is recorded with the value, so its origin and dependencies are always clear.

There is no “confidence score” for each value: an unverifiable percentage offers reassurance without proof. A source or line of reasoning can be checked; that is what your teams do, value by value, when supervising a batch.

Body material

chrome-plated brass

“Body material” row in the specifications table

🔗 data-sheet.pdf · p.4

Spray head

pull-out · 2 spray patterns

inferred from the product photo — labelled as such

🔗 product-photo.jpg

Worktop drilling

Ø 35 mm

inferred by reasoning: 35 mm cartridge + single-hole installation

↳ depends on: cartridge · installation type

// a growing priority

Your products in AI assistant answers

Buyers increasingly ask AI which product meets their needs. A complete, structured and accessible product page gives engines more material to understand the product and support their answer. GEO, or Generative Engine Optimisation, improves that clarity without guaranteeing a citation or recommendation on its own. Indexation, authority and query context also matter.

Clear limitations

What KaraK does not do

KaraK is not a PIM

It does not manage your publishing workflows: it fills your PIM — Akeneo or another platform — with proven data.

No shared database

KaraK does not impose a generic catalogue: it works with the products, sources, throughput and rules relevant to your market.

No magic confidence score

An enriched value does not carry an unverifiable “87% confidence” score: it has a reading explanation and a source — or is marked as inferred.

No automation without human oversight

AI proposes and human oversight validates: batches pass through human control points, and ambiguous cases are resolved rather than averaged.

Delivery

How does KaraK make your data available?

Through a dedicated interface with the required access, views and alerts, or directly in your information system. Akeneo integration is native, with a dedicated channel, parent/variant management and schema validation in a loop until data imports cleanly. KaraK also delivers through CSV, Excel, JSON, databases or APIs. Technical distribution standards — FAB-DIS and ETIM — are supported natively on input and output.

Dedicated interfaceUser accessDashboards · alertsAkeneo & other PIMsCSV · Excel · JSONDatabase · APIFAB-DIS · ETIM

Frequently asked questions

Questions we are often asked

Does KaraK replace a PIM?

No. KaraK is not a PIM and does not aim to become one: it is the layer that feeds it. Your product records remain managed in your tool — Akeneo or another PIM, an ERP, or simple files. KaraK researches, enriches, matches and checks the data, then delivers it into your catalogue with its sources.

How are sources selected?

Research agents apply a hierarchy of trust: manufacturer first, followed by retailers and distributors recognised for the quality of their product pages. Questionable sources — clearance sites, unverified marketplaces and unchecked specifications — are excluded. For every product, agents record what was retained and why.

How are contradictions between sources handled?

Through cross-checking and source hierarchy: manufacturer data takes precedence, values are compared across documents — product page, technical PDF and photo — and every written value includes a reading explanation stating where it was found and how it was interpreted. Anything not read directly is explicitly marked as inferred, whether from an indirect visual reading or from reasoning based on other established characteristics. Human oversight resolves ambiguous cases.

How are results delivered back into the information system?

Through your preferred channel: a dedicated interface with user access, dashboards and alerts; native Akeneo integration; CSV, Excel or JSON exports; a database or an API. Every delivered value retains its link to the source read by the agent — traceability travels with the data.

What volumes does KaraK process in production?

KaraK is in production on a catalogue of 640,000 products, with tens of thousands of products processed every year through a subscription. The matching module performed more than one million matching operations in one month. Batch orchestration — error recovery, quotas and oversight — is designed for this scale.

See KaraK work on your own products

Send us a sample of your catalogue: we will show you the enriched data, complete with sources — with no commitment.

Book a demo Build a product record live

KaraK also processes data beyond products — see our custom workflows