# Uthereal AI glossary Definitions of the sovereign AI, retrieval and enterprise knowledge terms used across uthereal.ai. Last updated: 2026-09-18. Source: https://uthereal.ai/glossary ## Sovereign AI Also known as: AI sovereignty, sovereign AI definition. Sovereign AI is artificial intelligence that runs entirely under one organisation's or country's legal and physical control: the models, the data, the logs and the infrastructure all stay inside a chosen jurisdiction, and no third party can access or retain the content. Sovereignty is not the same as encryption or a data-processing agreement. A system is only sovereign when the operator can prove where every byte is stored, which parties can technically reach it, and where inference actually happens. If answers are generated by an API in another country, the deployment is not sovereign, however strong the contract is. Regulated buyers ask for sovereignty when the content itself is the sensitive asset: legal files, patient records, defence material, unpublished research, government casework. For them the question is not whether a model is smart, but whether an audit can reconstruct exactly what left the building. At Uthereal: Uthereal Cortex runs on Swiss infrastructure or inside your own data centre, with an egress record that shows every outbound call - or the absence of one. Auf Deutsch - Souveräne KI: Souveräne KI bezeichnet KI-Systeme, die vollständig unter der rechtlichen und physischen Kontrolle einer Organisation oder eines Landes betrieben werden: Modelle, Daten, Protokolle und Infrastruktur verbleiben in der gewählten Jurisdiktion, ohne Zugriff Dritter. Related: Sovereign AI with Cortex (https://uthereal.ai/cortex/sovereign-ai), Swiss sovereign AI (https://uthereal.ai/swiss-sovereign-ai), AI for government (https://uthereal.ai/solutions/government) Permalink: https://uthereal.ai/glossary/sovereign-ai ## Data residency Also known as: AI data residency, where is my AI data stored. Data residency is the guarantee that a defined set of data is stored and processed in a named country or region. For AI systems it must cover four things: the source documents, the vector index, the prompt and answer logs, and the model inference itself. Most AI vendors answer residency questions with storage only. That leaves the largest exposure untouched, because prompts and retrieved passages are sent to whichever model endpoint the vendor uses, often in another jurisdiction, and are frequently retained for abuse monitoring. A credible residency statement names the data centre region for each of the four layers, states the retention period for logs, and lists every subprocessor that can technically read content. At Uthereal: With Uthereal, storage, index, logs and inference all sit in the same chosen region - Switzerland by default, or your own infrastructure. Auf Deutsch - Datenresidenz: Datenresidenz bedeutet, dass Daten in einem festgelegten Land gespeichert und verarbeitet werden. Bei KI-Systemen muss das vier Ebenen umfassen: Quelldokumente, Vektorindex, Prompt- und Antwortprotokolle sowie die Modellinferenz selbst. Related: Swiss private AI (https://uthereal.ai/swiss-private-ai), Security at Uthereal (https://uthereal.ai/trust/security), Private AI (https://uthereal.ai/solutions/private-ai) Permalink: https://uthereal.ai/glossary/data-residency ## Private AI Also known as: on-premise AI, self-hosted AI. Private AI is an AI deployment in which your content is never sent to a shared public model service and is never used for training. It runs in a single-tenant environment - your cloud account, your data centre, or a dedicated sovereign region - with isolated storage and isolated inference. The line that matters is tenancy. A public API with a no-training promise still means your prompts travel to shared infrastructure operated by someone else. A private deployment means the compute serving your answers serves only you. Private does not have to mean slow or old. Modern open-weight models run well on dedicated GPUs, and for most enterprise knowledge work the quality ceiling is set by retrieval quality, not by the frontier model. At Uthereal: Uthereal ships the same product in three shapes: Swiss-hosted, your cloud tenant, or fully on-premise with no outbound network path. Auf Deutsch - Private KI: Private KI bezeichnet KI-Systeme, bei denen Inhalte nie an öffentliche Modelldienste gesendet und nie für Training verwendet werden. Betrieb erfolgt in einer dedizierten Umgebung - eigene Cloud, eigenes Rechenzentrum oder souveräne Region. Related: Private AI (https://uthereal.ai/solutions/private-ai), Swiss private AI (https://uthereal.ai/swiss-private-ai), Secure knowledge base AI (https://uthereal.ai/solutions/secure-knowledge-base-ai) Permalink: https://uthereal.ai/glossary/private-ai ## RAG (retrieval-augmented generation) Also known as: retrieval augmented generation, what is RAG. RAG, or retrieval-augmented generation, is a pattern where a system first retrieves relevant passages from your own content and then asks a language model to answer using only those passages. The model supplies the language; your documents supply the facts. A RAG pipeline has four stages: ingest and split the source material, index it for retrieval, retrieve the best passages for a question, and generate an answer constrained to what was retrieved. Each stage can be the weak link, and in practice retrieval quality - not model choice - decides whether answers are trustworthy. RAG is preferred over fine-tuning for knowledge that changes, because you update a document instead of retraining a model, and because every claim can be traced back to a source. At Uthereal: Cortex is a managed RAG stack: ingestion, chunking, hybrid retrieval, reranking and cited generation, available through an API and the Uthereal SDK. Auf Deutsch - RAG (Retrieval-Augmented Generation): RAG bezeichnet ein Verfahren, bei dem zuerst passende Passagen aus den eigenen Inhalten gesucht und anschliessend von einem Sprachmodell zu einer Antwort verarbeitet werden. Die Fakten stammen aus den Dokumenten, nicht aus dem Modell. Related: Enterprise RAG chatbot (https://uthereal.ai/solutions/enterprise-rag-chatbot), RAG with citations (https://uthereal.ai/developers/rag-with-citations), How Cortex works (https://uthereal.ai/cortex/how-it-works) Permalink: https://uthereal.ai/glossary/rag ## Grounding Also known as: grounded answers, source-grounded AI. Grounding means constraining an AI answer to specific retrieved evidence, so that every statement can be traced to a passage in a known source. An ungrounded answer is generated from model memory alone and cannot be verified. Grounding is enforced, not requested. Strong systems pass only retrieved passages into the prompt, require the model to cite passage identifiers, and refuse to answer when retrieval returns nothing relevant. A refusal is a feature: it is the honest outcome when the knowledge base does not contain the answer. At Uthereal: Cortex answers carry passage-level citations with document, page and section, and return an explicit no-answer when the evidence is missing. Related: RAG with citations (https://uthereal.ai/developers/rag-with-citations), PDF page citations (https://uthereal.ai/developers/pdf-citations) Permalink: https://uthereal.ai/glossary/grounding ## Hallucination Also known as: AI hallucination, how to stop AI hallucinations. A hallucination is a confident AI statement that is not supported by any source. It happens because language models predict plausible text rather than look facts up, so gaps in retrieval get filled with fluent invention. Hallucination is reduced by engineering, not by prompt wording alone: better retrieval so the right passage is present, strict grounding so the model can only use what was retrieved, citation checks so unsupported sentences are caught, and an explicit refusal path when evidence is thin. Measure it. A small evaluation set of real questions with known answers tells you more about a deployment than any vendor benchmark. At Uthereal: Cortex reports the evidence behind each answer, so an unsupported claim is visible rather than hidden in fluent prose. Related: RAG with citations (https://uthereal.ai/developers/rag-with-citations), Cortex features (https://uthereal.ai/cortex/features) Permalink: https://uthereal.ai/glossary/hallucination ## Chunking Also known as: document chunking, chunk size RAG. Chunking is the process of splitting documents into passages small enough to retrieve precisely and large enough to stay meaningful. Chunk boundaries decide what an AI system can find, so poor chunking silently caps answer quality. Naive fixed-length splitting cuts tables in half and separates a heading from the rule it introduces. Structure-aware chunking follows the document's own layout - sections, clauses, table rows, slide frames - and keeps the parent context attached to each passage. Chunk metadata matters as much as chunk text: document title, section path, page number, effective date and access level all travel with the passage and make citation and permission filtering possible. At Uthereal: Uthereal chunks by document structure and keeps page and section metadata, which is what makes page-accurate citations possible. Related: Knowledge Object (https://uthereal.ai/atlas/knowledge-object), Chatbot over documents (https://uthereal.ai/developers/chatbot-over-documents) Permalink: https://uthereal.ai/glossary/chunking ## Embedding Also known as: vector embedding, text embeddings. An embedding is a numeric vector that represents the meaning of a piece of text, so passages with similar meaning sit close together in vector space. Embeddings let a system retrieve by meaning rather than by exact keyword match. Semantic search alone is not enough for enterprise content, where exact identifiers, product codes and legal references matter. Hybrid retrieval combines vector similarity with keyword search and then reranks the merged candidates. Embedding models are also a data-residency question: sending text to a hosted embedding API exports the same content you were trying to protect. At Uthereal: Cortex runs hybrid retrieval with reranking, and can run embedding inside your own environment when sovereignty is required. Related: Vector database alternative (https://uthereal.ai/developers/vector-database-alternative), Uthereal SDK (https://uthereal.ai/developers/sdk) Permalink: https://uthereal.ai/glossary/embedding ## Vector database Also known as: vector store, do I need a vector database. A vector database stores embeddings and returns the nearest matches for a query vector. It is one component of a retrieval system, not a complete one: it does not parse documents, enforce permissions, rerank results or produce citations. Teams often start by choosing a vector database and discover the real work sits around it - ingestion, structure-aware chunking, hybrid search, access control, evaluation and citation plumbing. That surrounding layer is where most build timelines are lost. At Uthereal: Uthereal replaces the assembly job: retrieval, permissions and citations arrive as one API instead of six components to integrate. Related: Vector database alternative (https://uthereal.ai/developers/vector-database-alternative), Uthereal SDK (https://uthereal.ai/developers/sdk) Permalink: https://uthereal.ai/glossary/vector-database ## Knowledge Object Also known as: knowledge object definition, structured knowledge unit. A Knowledge Object is a self-contained unit of expertise - the content, its structure, its provenance, its permissions and its validity period held together - so an AI system can use it, cite it and enforce rules about it. It is Uthereal's alternative to loose files and raw text chunks. Treating knowledge as objects rather than documents makes the hard questions answerable: who may see this, which version applies today, where did this claim come from, and what happens downstream when it is superseded. At Uthereal: Knowledge Objects are the unit Cortex ingests, governs, retrieves and licenses. Related: The Knowledge Object (https://uthereal.ai/atlas/knowledge-object), AI knowledge management (https://uthereal.ai/solutions/ai-knowledge-management) Permalink: https://uthereal.ai/glossary/knowledge-object ## Knowledge Agent Also known as: knowledge agent vs skills agent. A Knowledge Agent answers questions from a governed body of source material and cites the passages it used. Its job is to know: it retrieves, reasons over evidence and refuses when the evidence is missing, rather than taking actions in other systems. The counterpart is a Skills Agent, which executes multi-step expert workflows - drafting, checking, routing, filing. Most enterprise deployments need both, but they fail for different reasons and should be evaluated separately. At Uthereal: Cortex ships both agent types on the same governed knowledge layer, so a workflow agent inherits the same permissions and citations. Related: Knowledge vs Skills Agents (https://uthereal.ai/cortex/knowledge-agents-vs-skills-agents), Enterprise AI agents (https://uthereal.ai/cortex/enterprise-ai-agents) Permalink: https://uthereal.ai/glossary/knowledge-agent ## Context engineering Also known as: context window management, prompt context. Context engineering is the discipline of deciding what an AI model sees at the moment it answers: which passages, in what order, with which metadata and instructions. It has replaced prompt wording as the main lever on answer quality. Bigger context windows did not remove the problem; they changed it. Models attend unevenly across long inputs, so stuffing a whole handbook into the prompt degrades accuracy and cost at once. Selecting twelve good passages beats pasting two hundred pages. At Uthereal: Uthereal treats context assembly as a product surface: retrieval, reranking, deduplication and budget control are configurable rather than hidden. Related: Context is all you need (https://uthereal.ai/atlas/context-is-all-you-need), How Cortex works (https://uthereal.ai/cortex/how-it-works) Permalink: https://uthereal.ai/glossary/context-engineering ## White-label AI chatbot Also known as: white label chatbot, branded AI assistant. A white-label AI chatbot is an assistant you deploy under your own brand, domain and design, with no vendor badge and no vendor branding in the interface. The vendor supplies the engine; the customer owns the experience and the data. For publishers, associations and software vendors the chatbot is part of the product, so a visible third-party badge undermines it. The important checks are custom domain support, full theming, data isolation per tenant, and an export path if you ever leave. At Uthereal: Uthereal ships white-label assistants on your domain, Swiss-hosted, with no 'powered by' badge and typical launch in two to four weeks. Related: White-label AI chatbot (https://uthereal.ai/solutions/white-label-ai-chatbot), Enterprise RAG chatbot (https://uthereal.ai/solutions/enterprise-rag-chatbot) Permalink: https://uthereal.ai/glossary/white-label-ai-chatbot ## Permission-aware retrieval Also known as: access control RAG, document-level permissions AI. Permission-aware retrieval applies each user's access rights at query time, so the system can only retrieve passages that user is already allowed to read. Filtering after generation is too late: the content has already entered the answer. The common failure is an index built with an admin service account. Every user then effectively queries every document, and an innocuous question surfaces salary bands or an unannounced acquisition. Permissions must be attached to passages at ingest and enforced inside the retrieval query. At Uthereal: Cortex carries source permissions into the index and enforces them per request, with an audit trail of what was retrieved for whom. Related: Secure knowledge base AI (https://uthereal.ai/solutions/secure-knowledge-base-ai), Security (https://uthereal.ai/trust/security) Permalink: https://uthereal.ai/glossary/permission-aware-retrieval ## EU AI Act Also known as: AI Act compliance, EU AI regulation. The EU AI Act is the European Union's risk-based regulation of artificial intelligence. It classifies systems by risk level and imposes duties on providers and deployers, including transparency, human oversight, data governance and technical documentation for higher-risk uses. For internal knowledge assistants the practical burden is documentation and traceability: what data the system uses, who can access it, how outputs are produced, and how a person can review and override them. Systems that cite their sources and log their retrievals are far cheaper to document than opaque ones. Switzerland is not bound by the Act, but Swiss providers serving EU customers usually align with it, alongside the Swiss FADP. At Uthereal: Cortex produces the artefacts compliance teams ask for: source citations, retrieval logs, access records and an egress record. Auf Deutsch - EU AI Act: Der EU AI Act ist die risikobasierte KI-Verordnung der Europäischen Union. Je nach Risikoklasse gelten Pflichten zu Transparenz, menschlicher Aufsicht, Daten-Governance und technischer Dokumentation für Anbieter und Betreiber. Related: AI for government (https://uthereal.ai/solutions/government), Security (https://uthereal.ai/trust/security) Permalink: https://uthereal.ai/glossary/eu-ai-act ## Egress record Also known as: outbound data log, AI data egress. An egress record is a log of every outbound network call an AI deployment makes, including the destination, the payload class and the time - or a signed statement that no outbound call occurred. It turns 'your data stays here' from a promise into evidence. Auditors and security teams rarely dispute intentions; they ask for proof. An egress record is the cheapest proof to produce and the hardest to fake, which is why it is becoming standard in sovereign and government deployments. At Uthereal: Uthereal on-premise deployments ship an egress record by default, showing the absence of outbound traffic rather than asserting it. Related: AI for government (https://uthereal.ai/solutions/government), Sovereign AI (https://uthereal.ai/cortex/sovereign-ai) Permalink: https://uthereal.ai/glossary/egress-record