Delivered and demonstrable
Query your documents. Nothing leaves your infrastructure.
Knowledge AI answers natural-language questions about your document base. Every answer cites its source, down to the document, page and passage; where the information is not in your documents, the product says so rather than inventing an answer. Models, indexes and logs run within your own infrastructure, including in disconnected environments: your documents come in, and they stay inside.
What Knowledge AI does
Every capability below exists today and can be demonstrated on your own documents, in a working session.
Hybrid search
Search combines lexical search, on exact terms, with vector search, on semantic proximity: it finds the relevant passage even when your question uses different vocabulary. The evidence check that precedes the answer may still refuse a question phrased far from the text; a fix is in progress.
Verifiable citations
Every answer points to the document, page and passage it comes from. A colleague checks the source in one click, without taking the product at its word.
No answer without a source
When your documents do not contain the answer, the product says so and refrains. It does not fill the silence with a plausible-sounding answer.
Document versions
Successive revisions of a document are kept apart: the answer relies on the version in force and states which version it cites.
Role-based access control
Users only receive answers drawn from documents their rights give them access to. Rights are defined by your teams.
Audit log
Sign-ins and administrative actions are written to the audit log, which is read-only. Questions and answers are kept for thirty days in the conversations, which the administrator can consult; for each answer, an exportable trace gives the date, the account, the outcome, the corpus version and the sources cited. Everything resides within your infrastructure.
Reproducible quality evaluation
Answer quality is measured on a question set established in advance, including expected answers and expected refusals. A deterministic evaluation mode makes it possible to replay the measurement identically and to compare two versions on the same questions.
Offline operation
No internet connection is required.
Knowledge AI runs without any internet access: language models, indexes and logs reside on your servers. This capability, which products built on an external API cannot offer, opens it up to isolated environments, classified networks and industrial sites cut off from the outside world.
- No call to any external API, at any stage of processing
- Models run on your hardware, within your network
- Verifiable operation: the product answers with the network cable unplugged
A product that knows when to stay silent
An invented answer costs more than no answer at all. When the source is missing, Knowledge AI says so, and the evaluation report measures how often, and how rightly, it refuses.
Replayed from a real trial
Trial of 6 October 2026: asked five times, refused five times. The corpus consists of sixteen articles of the French Labour Code (Légifrance); the silence concerns this corpus, not French labour law. The trial was run in French.
The question
What does the company risk if the prevention plan has not been drawn up?
As asked, in French (the trial was run in French):
Que risque l'entreprise si le plan de prévention n'a pas été établi ?
It follows naturally from reading the obligations attached to the prevention plan, which the corpus sets out.
What the tool read
The six passages read deal with the prevention plan and risk assessment; none provides for a penalty when no prevention plan has been drawn up. The tool does not fill that silence: it stays silent.
View the full screenshotWhat you receive
The product, in your infrastructure
Installed on your servers, configured on your corpora and access rights, operated by your teams.
The evaluation report
Answer accuracy, correctness of refusals, quality of citations: measured on a question set agreed with you, together with the measurement context.
The code and documentation
You receive the code, documented installation and a skills transfer. You can evolve the product without us.
Design frameworks
ISO/IEC 42001
Knowledge AI is designed according to the requirements of ISO/IEC 42001, the standard for artificial intelligence management systems: governance, risk control, traceability.
OWASP
Knowledge AI follows OWASP recommendations for application security, including those for applications built on large language models.
Knowledge AI is not certified to ISO/IEC 42001; its design is aligned with the requirements of the standard.
AI Assessment
AI Assessment: starting from your own documents
Unsure which use cases to pursue? The AI Assessment examines your data and processes in a workshop. It results in a costed roadmap and a working prototype built on your own documents, without those documents ever leaving your perimeter.
Workshop
Your business and IT teams set out the intended uses, the data available and the security constraints.
Data analysis
Corpora and data flows are examined on site or within your environment: volume, formats, quality, sensitivity.
Roadmap and prototype
You receive a costed roadmap and a working prototype on your documents, to inform the decision.
See Knowledge AI answer on your documents
A demonstration is organised on a corpus you choose, within your environment if you wish. You judge on the evidence.