Code guarantees coverage.
Agents retain judgement.
This is KaraK's core architectural principle: everything that must be exhaustive — collecting documents, checking values and enforcing formats — is guaranteed mechanically, regardless of the AI model. The agent focuses on what requires judgement: reading, interpreting and choosing.
Specialised agents
One agent for each role — and they check one another
search agent
identifies the product and selects sources — manufacturer first, then recognised retailers
extraction agent
reads pages and documents, both text and visual renderings, and writes each value in the required format
vision agent
reads information that exists only in images: product photos, pictograms and dimensioned drawings
validation agent
checks every value against its source before saving it — verified, partial or not found
Separating roles is not cosmetic: the agent that validates is not the agent that extracted. This allows the workflow to correct itself — a value that cannot be found in its stated source returns to extraction with the exact reason, within the same session.
Batch orchestration
Designed for hundreds of thousands of products
At the scale of a real catalogue, the question is not “does it work?” but “what happens when something breaks?”. KaraK's orchestrator is designed around that question:
Multi-level quotas
Platform, customer and session: concurrency is capped at every level to share resources fairly, while a minimum interval between launches smooths the load.
A job queue with clear states
Every product becomes a prioritised job: queued, running, completed, failed or cancelled. The dashboard shows progress in real time.
Proactive pause and automatic recovery
When conditions deteriorate, the session pauses itself. Queued jobs are retained, never lost, and processing resumes when conditions improve.
Circuit breaker with a probe job
Too many consecutive errors? The session stops and periodically retests with a single probe job, spacing out attempts instead of wasting resources on an unresponsive website.
Stalled jobs detected
A process that remains stuck is automatically marked as failed and retried — nothing remains in limbo indefinitely.
Recovery after an incident
If the orchestrator restarts, in-progress tasks are recovered cleanly — a batch that has started is a batch that gets finished.
Agentic economics
The right model for the right task
Not every operation requires the same intelligence. A lightweight model can read a page and summarise it; deciding where to navigate, which links to open and when to stop — judgement — deserves a more capable model. Checking that a value follows a format needs no model at all: that is code. Using the largest model at every stage means paying for judgement when all you need is reading.
Read, summarise, extract
lightweight model — high throughput at a sustainable cost
Navigate, assess, decide
advanced model — judgement where it matters
Check, validate, format
code, not a model — deterministic and free
The intelligence of an agentic platform does not reside solely in its models: it also lies in routing. You cannot guess which stages can use a lightweight model without sacrificing quality — you measure it on benchmark products under consistent safeguards. This is where guardrails prove their value: once code guarantees coverage, a well-constrained mid-sized model can match the rigour of a large model for a fraction of the cost. Reliability comes from constraints, not model size.
The economics matter on both sides. This routing makes it sustainable to process hundreds of thousands of products — viable volume pricing for the customer and a viable operation for the software provider, without cutting quality. A provider that uses the largest model everywhere charges for its inefficiency; one that uses the smallest everywhere charges for its errors.
Oversight
AI proposes and executes.
Humans stay in control.
A real-time progress dashboard, clear states — active, paused, circuit breaker — with the reason displayed and manual start or stop whenever needed. Every workflow ends with systematic human validation: categorisation is approved, enrichment is reviewed and ambiguous cases are resolved. No batch is delivered without oversight.
Data sovereignty
Hosted in France
KaraK is hosted on French infrastructure, from ingestion to delivery. Your catalogues, taxonomies and rules remain yours: there is no shared database and your data is never reused for other customers.
See this architecture at work
The home-page demo shows the workflow in action — or, better still, see it working on your own products.