Glossary

You can refer here for definitions of key terms used throughout this workshop.

Term	Definition
Agent Sandbox	A secure execution environment for AI agents managed by the Agent Sandbox Controller. Provides pod scheduling with optional warmpool support for lower cold-start times. Combined with OpenShell’s application-level isolation and optionally Kata Containers for hardware-level isolation, it forms the production runtime for agents on Red Hat AI.
AI Agent	A software system that uses an LLM to reason about tasks, decide which tools to call, and take actions autonomously. Unlike a simple chatbot, an agent can chain multiple tool calls and LLM invocations to accomplish complex goals.
AutoML	Automated Machine Learning. A capability in RHOAI that lets you train and compare ML models through a guided wizard without writing model training code. You upload data, select a prediction type, and AutoML builds and ranks multiple models.
AutoRAG	Automated Retrieval-Augmented Generation. A feature in the RHOAI GenAI Playground that enables RAG with a simple toggle and file upload, without requiring manual vector database setup.
Chatterbox	A company acquired by Red Hat whose advanced risk analysis technology has been integrated into Garak, EvalHub, and SDG Hub. Enables a full adversarial testing flow: define a policy, generate synthetic attack prompts via SDG Hub, attack the model via Garak, and produce a report via EvalHub showing where the model breaks.
EvalHub	A centralized evaluation management capability in RHOAI that coordinates evaluation pipelines, stores results, and enables comparison across model versions and prompt iterations.
Garak	An open-source LLM vulnerability scanner that probes models for weaknesses including prompt injection, jailbreak attempts, data leakage, and bias. Used for pre-deployment adversarial testing.
Guardrails Orchestrator	A GA component in RHOAI that coordinates guardrail evaluation at the inference boundary. Works with NeMo Guardrails detectors (deployed as sidecars for lightweight features or via Hugging Face detector server for larger models) to provide input and output filtering.
GenAI Playground	An interactive environment in RHOAI for experimenting with deployed models. Supports system prompt configuration, temperature tuning, RAG knowledge attachment, and MCP tool connection without writing code.
LangGraph	An open-source framework for building stateful, multi-agent applications with LLMs. Fed Aura Capital uses LangGraph to orchestrate its five persona-based agents.
llm-d	A distributed inference engine that scales LLM serving across multiple replicas with features like prefix caching and smart routing. Available as a deployment resource option in RHOAI.
MaaS (Models as a Service)	A capability in RHOAI that publishes a deployed model through a managed gateway API with subscription tiers, usage metering, and API key authentication. Multiple consumers share a single model deployment.
MCP (Model Context Protocol)	An open standard for connecting AI agents to external tools and data sources. MCP servers expose tool capabilities that agents can discover and call at runtime.
MCPServerRegistration	A Kubernetes Custom Resource Definition (CRD) in RHOAI that registers MCP servers as first-class cluster resources, making them discoverable and manageable alongside other AI assets.
MLflow	An open-source platform for managing the ML lifecycle. In this workshop, MLflow provides tracing (recording agent decision paths), evaluation (measuring agent quality), and prompt registry (version-controlling system prompts).
Kata Containers	OpenShift sandboxed containers that provide hardware-level isolation by running each pod in a lightweight VM with its own kernel. Complementary to OpenShell’s application-level sandboxing in a defense-in-depth model: OpenShell constrains what the agent process can do within its container, Kata ensures that even if the agent escapes its container, it is still inside a VM boundary.
NemoGuardrails	NVIDIA’s open-source toolkit for adding programmable guardrails to LLM applications. Provides input shields (filtering harmful inputs) and output shields (filtering harmful outputs) at the inference boundary.
OCSF (Open Cybersecurity Schema Framework)	A standardized schema for security telemetry. OpenShell uses OCSF v1.7.0 for structured audit events, making them compatible with SIEM systems.
OGX (Open GenAI Stack)	An agentic API server (previously Llama Stack, renamed to ogx) that implements the OpenAI Responses API and related agentic APIs. Provides a conformant, self-hosted alternative to calling OpenAI directly. Any framework that speaks the OpenAI standard (LangChain, LangGraph, CrewAI) can use ogx as a drop-in backend without code changes. Also provides RAG via OpenAI-compatible vector store APIs across 14+ providers. GA targeted for RHOAI 3.5.
Policy Advisor	An audit2allow-style workflow in OpenShell for iterative policy discovery. When an agent’s action is blocked, the structured deny response tells the agent what was rejected and why. The agent can request access, a human reviews and approves in the CLI/TUI, and approved rules are hot-loaded into the running sandbox. Disabled by default to prevent over-permissioning.
Responses API	The strategic API for agentic workloads, based on OpenAI’s OpenResponses specification. Provides state management, connectors, file search, and tool calling in a standardized format. ogx provides a fully conformant implementation. Any framework using the OpenAI standard can use ogx as a drop-in, governed, on-premise backend.
OpenShell	An open-source agent runtime developed by NVIDIA and integrated into Red Hat AI. Provides agent sandboxing (seccomp, Landlock, network namespaces), deny-by-default security policies, credential injection, OCSF audit logging, and SPIFFE/SPIRE identity. Three components: CLI/SDK (developer interface), Gateway (control plane), Supervisor (PID 1 in every sandbox).
RAG (Retrieval-Augmented Generation)	A technique that enhances LLM responses by retrieving relevant information from external documents before generating an answer. Instead of relying solely on training data, the model’s response is grounded in current, organization-specific knowledge.
SPIFFE/SPIRE	SPIFFE (Secure Production Identity Framework for Everyone) is a standard for service identity. SPIRE is the reference implementation. Together they provide cryptographic workload identity without hardcoded keys - each agent gets a short-lived, scoped identity token.
Trace	A record of an entire request through a multi-agent system, showing every step: which agents were invoked, which tools were called, what inputs and outputs were exchanged, and how long each step took.
Span	A single unit of work within a trace. For example, one LLM invocation, one tool call, or one agent routing decision. Spans are nested to show the parent-child relationship of operations.
vLLM	An open-source, high-throughput LLM serving engine. RHOAI includes vLLM-based serving runtimes for NVIDIA GPU, Intel Gaudi, and IBM Spyre accelerators.

Term

Definition

Agent Sandbox

A secure execution environment for AI agents managed by the Agent Sandbox Controller. Provides pod scheduling with optional warmpool support for lower cold-start times. Combined with OpenShell’s application-level isolation and optionally Kata Containers for hardware-level isolation, it forms the production runtime for agents on Red Hat AI.

AI Agent

A software system that uses an LLM to reason about tasks, decide which tools to call, and take actions autonomously. Unlike a simple chatbot, an agent can chain multiple tool calls and LLM invocations to accomplish complex goals.

AutoML

Automated Machine Learning. A capability in RHOAI that lets you train and compare ML models through a guided wizard without writing model training code. You upload data, select a prediction type, and AutoML builds and ranks multiple models.

AutoRAG

Automated Retrieval-Augmented Generation. A feature in the RHOAI GenAI Playground that enables RAG with a simple toggle and file upload, without requiring manual vector database setup.

Chatterbox

A company acquired by Red Hat whose advanced risk analysis technology has been integrated into Garak, EvalHub, and SDG Hub. Enables a full adversarial testing flow: define a policy, generate synthetic attack prompts via SDG Hub, attack the model via Garak, and produce a report via EvalHub showing where the model breaks.

EvalHub

A centralized evaluation management capability in RHOAI that coordinates evaluation pipelines, stores results, and enables comparison across model versions and prompt iterations.

Garak

An open-source LLM vulnerability scanner that probes models for weaknesses including prompt injection, jailbreak attempts, data leakage, and bias. Used for pre-deployment adversarial testing.

Guardrails Orchestrator

A GA component in RHOAI that coordinates guardrail evaluation at the inference boundary. Works with NeMo Guardrails detectors (deployed as sidecars for lightweight features or via Hugging Face detector server for larger models) to provide input and output filtering.

GenAI Playground

An interactive environment in RHOAI for experimenting with deployed models. Supports system prompt configuration, temperature tuning, RAG knowledge attachment, and MCP tool connection without writing code.

LangGraph

An open-source framework for building stateful, multi-agent applications with LLMs. Fed Aura Capital uses LangGraph to orchestrate its five persona-based agents.

llm-d

A distributed inference engine that scales LLM serving across multiple replicas with features like prefix caching and smart routing. Available as a deployment resource option in RHOAI.

MaaS (Models as a Service)

A capability in RHOAI that publishes a deployed model through a managed gateway API with subscription tiers, usage metering, and API key authentication. Multiple consumers share a single model deployment.

MCP (Model Context Protocol)

An open standard for connecting AI agents to external tools and data sources. MCP servers expose tool capabilities that agents can discover and call at runtime.

MCPServerRegistration

A Kubernetes Custom Resource Definition (CRD) in RHOAI that registers MCP servers as first-class cluster resources, making them discoverable and manageable alongside other AI assets.

MLflow

An open-source platform for managing the ML lifecycle. In this workshop, MLflow provides tracing (recording agent decision paths), evaluation (measuring agent quality), and prompt registry (version-controlling system prompts).

Kata Containers

OpenShift sandboxed containers that provide hardware-level isolation by running each pod in a lightweight VM with its own kernel. Complementary to OpenShell’s application-level sandboxing in a defense-in-depth model: OpenShell constrains what the agent process can do within its container, Kata ensures that even if the agent escapes its container, it is still inside a VM boundary.

NemoGuardrails

NVIDIA’s open-source toolkit for adding programmable guardrails to LLM applications. Provides input shields (filtering harmful inputs) and output shields (filtering harmful outputs) at the inference boundary.

OCSF (Open Cybersecurity Schema Framework)

A standardized schema for security telemetry. OpenShell uses OCSF v1.7.0 for structured audit events, making them compatible with SIEM systems.

OGX (Open GenAI Stack)

An agentic API server (previously Llama Stack, renamed to ogx) that implements the OpenAI Responses API and related agentic APIs. Provides a conformant, self-hosted alternative to calling OpenAI directly. Any framework that speaks the OpenAI standard (LangChain, LangGraph, CrewAI) can use ogx as a drop-in backend without code changes. Also provides RAG via OpenAI-compatible vector store APIs across 14+ providers. GA targeted for RHOAI 3.5.

Policy Advisor

An audit2allow-style workflow in OpenShell for iterative policy discovery. When an agent’s action is blocked, the structured deny response tells the agent what was rejected and why. The agent can request access, a human reviews and approves in the CLI/TUI, and approved rules are hot-loaded into the running sandbox. Disabled by default to prevent over-permissioning.

Responses API

The strategic API for agentic workloads, based on OpenAI’s OpenResponses specification. Provides state management, connectors, file search, and tool calling in a standardized format. ogx provides a fully conformant implementation. Any framework using the OpenAI standard can use ogx as a drop-in, governed, on-premise backend.

OpenShell

An open-source agent runtime developed by NVIDIA and integrated into Red Hat AI. Provides agent sandboxing (seccomp, Landlock, network namespaces), deny-by-default security policies, credential injection, OCSF audit logging, and SPIFFE/SPIRE identity. Three components: CLI/SDK (developer interface), Gateway (control plane), Supervisor (PID 1 in every sandbox).

RAG (Retrieval-Augmented Generation)

A technique that enhances LLM responses by retrieving relevant information from external documents before generating an answer. Instead of relying solely on training data, the model’s response is grounded in current, organization-specific knowledge.

SPIFFE/SPIRE

SPIFFE (Secure Production Identity Framework for Everyone) is a standard for service identity. SPIRE is the reference implementation. Together they provide cryptographic workload identity without hardcoded keys - each agent gets a short-lived, scoped identity token.

Trace

A record of an entire request through a multi-agent system, showing every step: which agents were invoked, which tools were called, what inputs and outputs were exchanged, and how long each step took.

Span

A single unit of work within a trace. For example, one LLM invocation, one tool call, or one agent routing decision. Spans are nested to show the parent-child relationship of operations.

vLLM

An open-source, high-throughput LLM serving engine. RHOAI includes vLLM-based serving runtimes for NVIDIA GPU, Intel Gaudi, and IBM Spyre accelerators.