Glossary
The vocabulary,
without detours.
The terms that recur in an infrastructure, hosting or artificial intelligence project, defined in one usable sentence. No jargon meant to impress: these definitions serve decisions.
- FRANCE / MONACO / SWITZERLAND
- 24 DEFINITIONS
Sovereignty
- Digital sovereignty
- The ability of an organisation to decide for itself who accesses its data, where it is processed and under which rules. It does not reduce to the geographic location of servers: it depends above all on the law the operator is subject to and on the effective control exercised over access.
- Cloud Act
- A 2018 US law allowing United States authorities to require a provider subject to US law to hand over data it holds, including when that data is stored on servers outside US territory. The determining factor is the operator’s control, not the hardware’s location.
- GDPR
- The General Data Protection Regulation, in force since 25 May 2018 in the European Union. It governs any processing of data concerning individuals: defined purpose, legal basis, limited retention, information of data subjects and exercise of their rights. Source: CNIL
- SecNumCloud
- A qualification framework published by the French cybersecurity agency (ANSSI) for cloud computing providers. It requires reinforced security guarantees and protection against unauthorised external access, including access based on non-European law. Source: ANSSI
- Health data hosting (HDS)
- A mandatory French certification for any host storing personal health data on behalf of a third party. It conditions the hosting of medical records, test results and identifying clinical research data.
Infrastructure
- On-premise
- A deployment mode in which servers and software are installed on the organisation’s own premises, on its own network and under its direct administration, rather than leased from a remote provider. Data never leaves the organisation’s physical perimeter.
- Private hosting
- Providing servers dedicated to a single organisation, as opposed to shared hosting or a cloud infrastructure shared between several clients. Resources are not shared and isolation does not depend on software configuration.
- Reversibility
- The ability to leave a provider or a technology without data loss or major interruption, within an acceptable time and cost. It is prepared from the design stage: open formats, up-to-date documentation, no dependency on a proprietary component.
- RTO (recovery time)
- The maximum acceptable duration of a service interruption before recovery becomes critical for the organisation. RTO determines the fallback architecture: an RTO of minutes requires immediate redundancy, an RTO of days allows rebuilding.
- RPO (accepted data loss)
- The volume of data an organisation accepts losing in case of incident, expressed as a duration. An RPO of one hour means data produced in the hour before the failure may be lost; it conditions backup frequency and replication mode.
- 3-2-1 backup
- A backup rule of keeping three copies of data, on at least two distinct technological media, including one copy kept off site. It protects simultaneously against hardware failure, human error, ransomware and a disaster hitting an entire site.
- Immutable backup
- A backup that cannot be modified or deleted for a defined period, including by an administrator holding all rights. It is the main protection against ransomware, which precisely seeks to destroy backups before encrypting data.
- Virtualisation
- A technique of running several isolated operating systems on the same physical machine, each behaving as a distinct server. It simplifies backups, speeds up restores and consolidates an underused server fleet.
- Network segmentation
- Partitioning a network into distinct zones, connected only by strictly necessary flows. It limits the spread of an intrusion or ransomware: a compromised workstation cannot directly reach database servers in another zone.
Artificial intelligence
- Local artificial intelligence
- Running artificial intelligence models on machines controlled by the organisation, inside its network, without calling a service hosted by a third party. The documents analysed and the questions asked never leave the organisation’s perimeter.
- RAG (retrieval-augmented generation)
- A method of first retrieving the relevant passages from the organisation’s documents, then handing those passages to a language model to draft an answer grounded in them. It reduces invented answers and allows the sources used to be cited.
- Hallucination
- The production, by a language model, of false information stated with confidence. The phenomenon is inherent to how probabilistic models work; it is reduced through documentary grounding, required source quotation and human validation of sensitive decisions.
- Open model
- An artificial intelligence model whose weights are published and can be freely downloaded and run on one’s own infrastructure, as opposed to a model available only through a remote interface. It enables offline use with no data transfer.
- Augmented document search
- Search combining classic document indexing with a language model, so as to answer a question phrased in everyday language rather than keywords. It returns the original passages, allowing the answer to be verified at its source.
Data
- Cognitive heritage
- The memory, knowledge, experience and know-how of a person or an organisation, considered as capital to be gathered, structured, protected and transmitted. The term designates both the material and the preservation approach.
- Intergenerational transmission
- The organised transfer of knowledge, memory or heritage from one generation to the next, beyond mere legal succession. For knowledge, it requires gathering, structuring and shaping work meant to stay understandable to people who never knew its author.
- Professional secrecy
- A legally required, criminally sanctioned duty not to disclose confidential information learned in the course of one’s profession. It concerns legal, medical and banking professions in particular, and rules out any processing relying on an unsupervised third-party service.
- Sub-processor
- Under the GDPR, a person or body processing personal data on behalf of a controller, on the controller’s instructions. A host or email provider is a sub-processor; it must offer sufficient guarantees and act only on documented instructions.
- Retention period
- The period during which personal data may lawfully be kept, set in advance according to the purpose pursued and the applicable legal obligations. At its end, data must be deleted or irreversibly anonymised.
A term missing, or a definition that seems imprecise? Let us know: this glossary is updated as questions arise on engagements.
Prendre contact
A technical question behind a glossary word?
Describe your environment: we answer on substance, with no jargon and no catalogue.