Foundational Reading · 67 Terms · Reviewed June 2026

Enterprise AI Glossary

Plain-English definitions for the terms enterprises actually use when buying, building, and governing on-premise AI. Covers AI agents, multi-agent orchestration, the EU AI Act, private RAG, LLM routing, agentic design patterns, and the operational vocabulary in between. Each term links to the relevant VDF AI pillar, blog post, or product page for deeper context.

Browse the Knowledge Hub Read the blog

A

# Agent Memory The mechanisms by which an AI agent retains and recalls information across interactions and sessions.

Agent memory spans short-term (within a conversation), long-term (across sessions), and episodic (specific past events). Without memory, agents repeat mistakes, lose context, and cannot learn from prior interactions. Enterprise memory adds access controls so agents only recall what they are authorised to. See Agent Memory and Memory Patterns for AI Agents.

B

# Bias Detection & Fairness Auditing Testing AI outputs for systematic group-level disparity and documenting mitigations.

Bias detection evaluates whether a model or workflow produces materially different outcomes across protected groups. It is required for high-risk EU AI Act systems and increasingly expected by procurement teams. See Bias Detection & Fairness Auditing.

Full definition: Bias Detection & Fairness Auditing →

C

# Chunking Splitting documents into smaller segments for embedding and retrieval in a RAG pipeline.

Chunking is the preprocessing step that determines RAG quality. Too large and the embedding is diluted; too small and you lose context. Strategies include fixed-size, sentence-boundary, semantic, and hierarchical chunking. The right choice depends on document type, embedding model, and downstream task. See RAG Technology Best Practices.

Full definition: Chunking →

# Conformity Assessment The EU AI Act procedure by which a high-risk AI system is shown to meet legal requirements before placement on the market.

A conformity assessment verifies that a high-risk AI system has a risk management system, technical documentation, data governance, logging, transparency, human oversight, accuracy and cybersecurity controls. The architecture decisions you make early — especially logging and traceability — determine whether this is straightforward or painful. See EU AI Act-Ready On-Premises AI Architecture.

Full definition: Conformity Assessment →

# Context Engineering Designing and managing the information an AI model receives so it produces the right output.

Context engineering is the discipline of controlling what goes into a model's context window: system instructions, retrieved passages, tool outputs, conversation history, and memory. It is more systematic than prompt engineering and critical for agent systems where context is assembled programmatically. See Context Engineering.

D

# Data Sovereignty Keeping data and the systems that process it inside a defined legal or geographic boundary.

Data sovereignty is why many regulated organizations cannot route prompts, retrieved passages, or logs through uncontrolled third-party services. It is the practical driver behind on-prem and sovereign-cloud AI architectures. See Data Sovereignty vs Data Residency and Sovereign AI with VDF AI.

Full definition: Data Sovereignty →

# Decision Traceability The ability to reconstruct, after the fact, why an agent or model produced a given output.

Decision traceability ties outputs back to the inputs, retrieved sources, tool calls, and model selection that produced them. Without it, “why did the system decide that?” has no answer. See Decision Traceability for Audits.

Full definition: Decision Traceability →

# DORA (Digital Operational Resilience Act) EU regulation requiring ICT resilience, incident reporting, and third-party risk controls for financial entities.

DORA brings AI vendors used by financial entities into the ICT third-party register, expects aligned incident classification, and requires resilience testing that includes AI workflows. Concentration on a single AI provider is a flagged risk, which is one reason regulated buyers prefer platforms with deployment flexibility. See AI Governance Framework for Regulated Industries.

E

# Embeddings Numeric vector representations of text (or other content) used for semantic retrieval.

Embeddings turn passages into vectors that can be searched by meaning rather than keyword. In a Private RAG architecture, the embedding store is sensitive — it contains a representation of internal documents — and is one of the first places enterprises ask “where does this live?”. See Embeddings and Private RAG.

Full definition: Embeddings →

# Energy-Aware Routing (Green AI) Routing decisions that consider the energy and carbon profile of each model, not only quality and cost.

Energy-aware routing prefers smaller, more efficient models for high-volume work where the quality threshold is met, reserving frontier models for steps that genuinely need them. See Orchestration & Energy research and the white paper How We Reduce Energy Consumption.

F

# Fine-Tuning Adapting a pre-trained model to a specific domain or task by training on curated examples.

Fine-tuning adjusts model weights to improve performance on domain-specific tasks — legal language, medical terminology, internal process steps — without training from scratch. In enterprises, the decision between fine-tuning, RAG, and routing is an architecture choice with cost, latency, and data-handling implications. See Fine-Tuning vs Routing vs Smaller Models and Model Fine-Tuning.

Full definition: Fine-Tuning →

# Foundation Model A large model trained on broad data that is adapted (via prompting, fine-tuning, or RAG) for downstream tasks.

Foundation models are the generic substrate of modern AI. Enterprises rarely use them raw; they layer retrieval, routing, and governance on top. For high-volume internal tasks, a smaller specialised model often outperforms a generic foundation model on cost, latency, and energy. See Small Language Models in Enterprise AI.

Full definition: Foundation Model →

G

# Governance Plane The layer of the AI stack that enforces policy across identity, models, tools, retrieval, and logs.

A governance plane is what turns a collection of agents into a controllable enterprise system. It is not a reporting dashboard added later; it is the runtime that decides who can do what, with which model, on which data. See AI Agent Governance.

Full definition: Governance Plane →

# GPAI Model General-Purpose AI Model — a model with broad capability across tasks, as defined in the EU AI Act.

GPAI providers face specific transparency, documentation, and (for systemic-risk models) evaluation and incident-reporting obligations. Deployers of GPAI inside a high-risk product inherit additional duties. See EU AI Act-Ready Architecture.

Full definition: GPAI Model →

# GraphRAG (Knowledge Graph RAG) Retrieval pattern that uses a knowledge graph as a structured backbone, often alongside vector search.

GraphRAG (also called Knowledge Graph RAG) treats entities, relationships, and properties as first-class retrieval targets. It outperforms pure vector retrieval on multi-hop and relational questions — policy applicability, contract obligations, supply chain dependencies — and combines well with vector retrieval in hybrid setups. See Knowledge Graph RAG.

H

# Hallucination A confident-sounding output that is not grounded in the retrieved or provided evidence.

Hallucinations are reduced — not eliminated — by good retrieval, citations, and evaluation. In an enterprise setting, the more dangerous failure mode is a hallucinated action, not a hallucinated answer: an agent that fabricates a justification for the wrong tool call. See Private RAG vs Enterprise Search and Agent Evaluation.

Full definition: Hallucination →

# High-Risk AI System An EU AI Act category covering AI used in safety components or in listed high-impact domains.

High-risk AI systems include those used in critical infrastructure, education, employment, essential services, law enforcement, migration, and the administration of justice. They trigger the bulk of EU AI Act obligations. Architecture matters because logging, oversight, and documentation cannot be retrofitted cheaply. See EU AI Act-Ready Architecture.

Full definition: High-Risk AI System →

# Human Oversight The requirement and capability for qualified humans to monitor, intervene in, and override AI system decisions.

Human oversight is a core EU AI Act requirement for high-risk systems. It ranges from simple approval gates to real-time monitoring dashboards that flag anomalous agent behaviour. The key question is not whether humans can intervene, but whether the system makes intervention practical and timely. See Human Oversight Requirements and Human-in-the-Loop.

Full definition: Human Oversight →

# Hybrid Deployment Running sensitive workloads on local infrastructure while selectively using cloud models for approved tasks.

Hybrid deployment is the realistic shape of most regulated enterprise AI. Retrieval, embeddings, logs, and routing stay local; specific workflows can call external models when policy allows and quality benefits justify it. See On-Premise AI Agent Platform.

Full definition: Hybrid Deployment →

K

# Knowledge Graph A structured representation of entities, relationships, and properties that machines can traverse and reason over.

A knowledge graph encodes facts as triples (entity–relationship–entity) and enables multi-hop queries that pure vector search cannot answer reliably. In enterprise AI, knowledge graphs power policy applicability checks, supply-chain tracing, and contract-clause navigation. See Knowledge Graph RAG and Federated Knowledge Graph Playbook.

Full definition: Knowledge Graph →

L

# LLM Inference The process of running a trained model to generate outputs from inputs — the runtime cost centre of AI.

LLM inference is where compute budget is spent. Latency, throughput, and cost per token depend on model size, quantisation, batching strategy, and hardware. Enterprises that run inference locally (via vLLM, Ollama, llama.cpp) trade cloud convenience for control and predictable economics. See LLM Inference and Where vLLM Fits.

M

# MCP (Model Context Protocol) An open protocol for exposing tools, data sources, and prompts to AI agents in a standard way.

MCP lets you wrap an internal API or a third-party agent platform as a tool any compatible agent can invoke. In practice it is the integration glue between platforms — for example, running n8n or CrewAI flows as tools inside a VDF AI Network. See Model Context Protocol and Integrate n8n & CrewAI as MCP Tools.

Full definition: MCP (Model Context Protocol) →

# Model Card Short structured documentation of a model: intended use, training data summary, evaluations, and limits.

A model card is the artefact procurement and compliance teams ask for. It is also useful internally: when routing decides which model handles a step, the card explains why that model is approved for that workload. See Model Evaluation Suite.

O

# Observability (AI Agent) Live and historical visibility into agent execution — what ran, on what data, with what result.

AI agent observability goes beyond LLM call logs. It needs to capture orchestration paths, tool invocations, approvals, retries, and downstream effects. See Agent Observability and AI Agent Observability: Logs, Traces, Audit.

Full definition: Observability (AI Agent) →

# On-Premise AI Agent Platform The operating layer for running agents, retrieval, routing, and governance inside controlled infrastructure.

An on-premise AI agent platform is more than a model host. It bundles agent runtime, private retrieval, policy enforcement, observability, and deployment control so the enterprise retains data and runtime control. See the pillar page, On-Prem AI overview, and Implementation Roadmap.

Full definition: On-Premise AI Agent Platform →

P

# Permission-Aware Retrieval RAG that respects source-system access controls instead of surfacing content the user cannot read.

Without permission-aware retrieval, the RAG layer becomes a permissions bypass: it indexes content the user would not be allowed to open in the source system and serves it back through the chat surface. See Private RAG.

Full definition: Permission-Aware Retrieval →

# Private RAG Retrieval-augmented generation over enterprise-controlled data, storage, and embeddings.

Private RAG differs from cloud RAG in three ways: where the embeddings live, who controls retrieval policy, and whether retrieval is audited as part of the AI platform. See the Private RAG pillar and Private RAG vs Enterprise Search.

Full definition: Private RAG →

# Prompt Injection An attack where malicious input causes an AI system to ignore its instructions or execute unintended actions.

Prompt injection — both direct (user input) and indirect (via retrieved content) — is the most discussed LLM security risk. Defences include input sanitisation, instruction-data separation, output filtering, and treating retrieved text as data rather than instructions. No single defence is complete; layered guardrails at the platform level are required. See AI Guardrails and Enterprise AI Agent Security.

Full definition: Prompt Injection →

# Provider vs Deployer EU AI Act roles: the provider places an AI system on the market; the deployer uses it under their authority.

Most enterprises are deployers, not providers. Deployer duties (oversight, monitoring, fundamental-rights impact assessment) are different from provider duties (conformity assessment, technical documentation, post-market monitoring). Knowing which role you play for each system is the start of any compliance plan. See EU AI Act-Ready Architecture.

Full definition: Provider vs Deployer →

R

# RAG (Retrieval-Augmented Generation) Retrieving relevant passages and passing them into a generation step so answers are grounded in evidence.

RAG reduces hallucination and gives users source citations. Quality depends on ingestion, chunking, embeddings, ranking, and freshness — “upload documents and chat” is the demo, not the architecture. See RAG Technology Best Practices.

S

# SEEMR (Self-Evolving Model Router) VDF AI's architecture for governed, observable, adaptive LLM routing across providers and local models.

SEEMR formalises model selection as a runtime decision shaped by policy, cost, latency, energy, and outcome feedback — rather than a static configuration. See the SEEMR architecture page and the white paper The Self-Evolving Model Router.

Full definition: SEEMR (Self-Evolving Model Router) →

# Semantic Search Search that matches by meaning rather than exact keywords, using vector similarity over embeddings.

Semantic search finds relevant content even when the query uses different words than the source document. It is the retrieval engine behind RAG and the reason AI assistants can answer questions about internal knowledge bases without exact keyword matches. See Semantic Search.

Full definition: Semantic Search →

# Shadow AI AI tools used inside the organization without security, procurement, or compliance approval.

Shadow AI is the gap between employees adopting new tools and IT discovering they exist. Recent surveys put it above 30% of weekly AI usage in many enterprises. You cannot govern what you cannot see. See AI Inventory & Shadow AI Discovery.

Full definition: Shadow AI →

# Small Language Model (SLM) A smaller, often specialised model that runs locally and trades general capability for cost, latency, and control.

Small language models are how enterprises make high-volume internal AI economically viable: many routine tasks (classification, extraction, summarization) do not need a frontier model. See Small Language Models in Enterprise AI.

Full definition: Small Language Model (SLM) →

# Sovereign AI AI operated under a specific jurisdiction's legal, infrastructure, and data-residency requirements.

Sovereign AI is the practical answer to “our data cannot leave the country / cannot leave our control.” It usually combines on-prem or sovereign-cloud deployment, private retrieval, and approved local or jurisdiction-bound models. See the AI Agent Security & Data Sovereignty pillar for the zero-trust architecture, and Sovereign AI with VDF AI for the product perspective.

Full definition: Sovereign AI →

T

# Tool Governance Scoping which agents may call which tools, under which conditions, with which approval steps.

Tool governance is where governance becomes operational. An agent that can read documents is one risk profile; one that can modify records or send customer-facing content is another. The orchestrator must enforce scope, not the calling code. See AI Agent Governance.

Full definition: Tool Governance →

# Tool Use / Function Calling The mechanism by which an AI model requests execution of an external function — an API call, database query, or code execution.

Tool use (also called function calling) is what turns a text generator into an agent. The model outputs a structured request (tool name + arguments), the runtime executes it, and the result is fed back. Enterprise tool use adds permission checks, rate limits, and audit logging around each invocation. See Tool Use & Function Calling and Tool Calling Patterns.

Full definition: Tool Use / Function Calling →

# Transparency Obligations EU AI Act duties to disclose AI use to affected people — for example, when interacting with a chatbot or seeing AI-generated content.

Transparency obligations apply broadly, not only to high-risk systems. Deployers of emotion-recognition or biometric-categorisation systems have additional disclosure duties; providers of generative AI must mark synthetic outputs. See AI Governance & Compliance Problems.

Full definition: Transparency Obligations →

V

# Vector Database A database optimized for similarity search over embedding vectors.

A vector database is the storage layer behind RAG. In a private deployment, it sits next to the documents it indexes — not in a separate vendor tenant — so retrieval, permissions, and audit live in one boundary. See Vector Database and Private RAG.

Z

# Zero Data Retention A processing mode in which the model provider does not store inputs or outputs after a request completes.

Zero data retention is one of the controls procurement teams ask for when sensitive prompts must traverse an external API. It is not a substitute for keeping data local — logs, embeddings, and retrieved passages still live somewhere — but it constrains what a third party can keep. See Why Data Security Matters in AI.

Full definition: Zero Data Retention →

# Zero-Trust AI Agent Architecture where every agent has its own identity, every tool call is authorized, and every output is logged.

Zero-trust applied to AI agents means no agent, model output, or retrieved passage is trusted by default. Identity is per-agent, tool access is least-privilege, retrieved text is treated as data rather than instructions, and execution traces are exportable for replay. This is the baseline for regulated and sovereign deployments. See AI Agent Security & Data Sovereignty.

Translating these terms into an architecture?

Book a working session with the VDF AI team. We will map your use cases to the right combination of on-premise platform, governance, retrieval, and routing.

Talk to VDF AI

Enterprise AI Glossary

A

B

C

D

E

F

G

H

K

L

M

O

P

R

S

T

V

Z

Go from terminology to architecture

On-Premise AI Agent Platform

AI Agent Orchestration

AI Agent Governance

Private RAG

LLM Routing

VDF AI Compliance

Translating these terms into an architecture?

Request a Demo

Thank You!