These figures are the NCE architectural targets — built on 6 months of development and real-sector testing. Not yet production results, because there is no production yet. Said clearly.
Defined from the outset — measured and refined on each deployment.
Not 18 months of IT project. Twelve weeks to go from signed brief to first results on your real use case. Before/after measurement, executive committee presentation in week 12.
RAG approaches plateau between 35 and 45% error on dense regulatory corpus. The KG + NCE architecture structurally targets the operational threshold of < 5% — through 14 reliability dimensions continuously measured.
No 18-month IT project before the first result. The NCE infrastructure — knowledge graph, vector engine, interface — is deployed in your infrastructure in 3 business days. Connection to your existing document sources included.
100% on-premise. No call to an external service. No cloud dependency. The engine, the network, the interface and the models all live within your perimeter — physical, virtualised or accredited enclave.
The < 5% score is the structural target of the NCE architecture — not a marketing promise. It is carried by 14 reliability dimensions continuously measured, an entity whitelist, out-of-domain refusal, and fidelity validation of every answer.
Measured on real phytosanitary corpus (2023–2024) · formal qualification in progress
Reference question sets with expected answers, replayed at each NCE version on the same corpus. No cherry-picking — same scope, same evaluators.
Every version is compared to the previous one on the same questions. KPIs: average score, pipeline confidence, refusal rate, contradictions detected. No figures published without this protocol.
All test sets are documented. Provided under NDA for clients in evaluation. The measurement methodology is part of the delivery contract.
Six months of work. A single validation ground: real regulatory archives, in the phytosanitary domain. These targets don't come from generic benchmarks — they come from documents you know.
In regulated industries, years of critical decisions sit side by side with no link between them — studies, assessment reports, internal notes — scattered across formats and servers. Regulatory expertise lives in people's heads, not in systems. When the expert leaves, the memory leaves with them. That is the precise problem, documented in the field, that NCE is built to solve.
Before a single line of code, one absolute rule: no regulatory data ever leaves the client's infrastructure. Ever. From these principles came the NCE engine — Neural Context Engine — designed to reason over your archives without ever exposing them, in a fully isolated environment with no external connection required. A non-negotiable requirement, dictated by the reality of the pharmaceutical, phytosanitary and nuclear industries.
NCE ingests your documents in every existing format — and does more than index them. It rebuilds the connections between the information it finds: which study covers which substance, which decision applies to which use, which standard takes precedence over which. Tested from the outset on real phytosanitary archives (marketing authorisations, ANSES). No answer without an identified source.
NCE integrates the domain's official reference data as an anchor of truth. The reliability of each answer is assessed against 11 measured criteria — grounded in established scientific studies, not arbitrary parameters. If an answer goes beyond the limits of the corpus, NCE refuses to answer rather than invent. In a regulatory context, it is that refusal that lends credibility to everything else.
NCE automatically checks whether each answer is faithful to the source documents — and flags it explicitly. It retrieves information even if you use a different term from the original author's. The test procedures cover 778 scenarios on real corpora. This isn't presentation — it's the rigour of a system built to withstand inspection.
The tests run on real document corpora to confirm the expected precision levels. Every measured gap sharpens the evaluation protocol — rigorous, documented work carried out in the field before any wider generalisation.
NCE combines three complementary search modes across your corpus to cover every angle of a regulatory question. Each answer's reliability is measured against 14 distinct criteria — and the answer is held back if it fails any of them. The first version able to question its own answer before delivering it to you.
NCE no longer just answers: it automatically spots and flags inconsistencies in your documents — a value out of range, outdated information, sources that contradict each other — with every alert justified and tied to its source. Where a conventional engine lets it through, NCE raises the flag.
NCE only admits into its memory information that is verified and tied to a recognised source; the rest is set aside. Less noise, a sound base that every answer can rely on with confidence.
NCE opens up to other regulated sectors beyond phytosanitary, adapts the level of its answer to each user's role, and extends its command of languages with an advanced multilingual mode. Also on the horizon: autonomous investigations run across your corpus.
No generic promises. 30-minute diagnostic: we assess what KnowWeave can do on YOUR data.