kroons is a single-tenant, retrieval-augmented knowledge platform that turns terabyte-scale document corpora into a citable answer engine — deployed inside your own infrastructure, governed by your own access policies.
Bearings on the centrifugal pumps in this project are replaced every 8,000 operating hours, with torque specified at 45 Nm. The figure is drawn directly from the maintenance datasheet…
[source: pump_spec_v3.pdf, p.12]The platform
A purpose-built stack that combines hybrid retrieval, contextual indexing and large language models — without ever surrendering your data or trusting a black box.
Dense semantic vectors and keyword search run in parallel, fused and neurally re-ranked so the most relevant passage surfaces — by meaning, not just matching words.
Sources are injected by the system, not invented by the model. Every answer carries a verifiable file-and-page reference you can open in one click.
Language-agnostic embeddings let users query in their own language across corpora written in another — Italian questions over English standards, and beyond.
A two-pass pipeline extracts document structure, then enriches every passage with section-level context before indexing — dramatically sharpening retrieval on dense technical material.
Project-scoped access control means users only ever retrieve what they're cleared to see — enforced at the data layer, not painted over in the UI.
Beyond documents, a natural-language agent queries your relational project databases — cost sheets, equipment registers, inspection logs — under the same interface.
The pipeline
An operator-supervised pipeline that turns an unstructured document store into a precise, governed retrieval index.
Source documents are scanned, diffed against the index, and queued for operator review before anything changes.
A two-pass pass extracts hierarchy and section context, then enriches each passage for unambiguous retrieval.
Contextualised passages are vectorised into a high-performance index alongside a parallel keyword index.
Hybrid search, fusion and re-ranking feed a language model that answers strictly from the retrieved evidence.
The architecture
Cleanly separated layers — each independently deployable and scalable — with your documents and indexes held entirely on-premise.
Security & governance
The platform is built to keep answers anchored to your evidence and your data inside your perimeter. If the documents don't support an answer, it says so rather than inventing one.
We're preparing the platform for wider deployment. Reach out for a technical walkthrough, a tailored evaluation, or to discuss early access.
Contact us →