Skip to content

Glossary

Common terms used throughout the AI Agents Learning Hub, organized alphabetically.

A

Agent A software system that perceives its environment, makes decisions, and takes actions to achieve a goal. AI agents use LLMs as a reasoning engine combined with tools, memory, and planning.

AgentGroupChat A Semantic Kernel construct that enables multiple agents to collaborate in a shared conversation. Each agent sees the conversation history and can respond, delegate, or call tools.

Agentic RAG An extension of basic RAG where an agent decides when to retrieve information, what query to use, and whether to retrieve again if the first result is insufficient.

A2A (Agent-to-Agent Protocol) A protocol pattern for letting independent agents discover each other's capabilities, delegate work, and exchange task results across system boundaries.

AG-UI A protocol for connecting agents to user interfaces so agent state, tool calls, messages, and UI events can stay synchronized.

ARM Template Azure Resource Manager template — a JSON file that declaratively describes Azure infrastructure. Can be deployed with the "Deploy to Azure" button.

B

Bicep A domain-specific language (DSL) for Azure infrastructure as code. Compiles to ARM JSON. Cleaner syntax than raw JSON templates. Used in the infra/ folder of this repo.

Chunking The process of splitting a large document into smaller pieces before embedding and indexing. Chunk size affects retrieval quality — too small loses context, too large loses precision.

C

Chat Participant A VS Code extension concept — a custom @agent that plugs into GitHub Copilot Chat. Registered via contributes.chatParticipants in package.json.

Completion A response from an LLM given an input prompt. Also called a "generation" or "inference."

Copilot Extension A GitHub integration that adds a custom agent to GitHub Copilot across github.com, VS Code, and other surfaces. Requires an API endpoint (webhooks).

Context Window The maximum amount of text (measured in tokens) an LLM can process in a single request. As of 2025, ranges from 8k (small models) to 1M+ tokens (Gemini 1.5).

Context Caching Reusing previously processed prompt context so repeated long-context calls can reduce latency and token cost.

D

Dense Embedding A numerical vector representation of text, where semantically similar texts have similar vectors. Produced by embedding models like text-embedding-3-small.

Document Grounding Providing retrieved documents to the LLM as context, so the answer is based on specific sources rather than the model's training data. Reduces hallucination.

E

Embedding A fixed-length numerical vector that captures the semantic meaning of text. See Lab 007 — What are Embeddings?.

Embedding Model A model that converts text to embeddings. Examples: text-embedding-3-small (OpenAI), text-embedding-ada-002, nomic-embed-text (local/Ollama).

Evaluation The process of measuring an agent or model against expected behavior, usually with test cases, rubrics, metrics, and regression checks.

F

Function Calling An LLM feature where the model can request to call a specific function with structured arguments, rather than returning plain text. Also called "tool use."

Foundry Agent Service Microsoft's managed service for deploying AI agents with built-in tool integration, tracing, and evaluation. Part of Azure AI Foundry.

G

GitHub Models A free service (github.com/marketplace/models) providing API access to leading LLMs (GPT-4o, Llama 3, etc.) using a GitHub token. No credit card required.

GraphRAG A RAG pattern that builds a knowledge graph of entities and relationships so retrieval can answer questions that require cross-document synthesis.

Grounding See Document Grounding.

Guardrails Runtime constraints that inspect or modify agent inputs, tool calls, and outputs to enforce safety, scope, privacy, or policy requirements.

H

Hallucination When an LLM generates confident-sounding but factually incorrect output. Grounding with RAG reduces hallucination by providing accurate source documents.

Hybrid Search Combining dense vector search (semantic similarity) with sparse keyword search (BM25/full-text) for better retrieval. Usually scored with Reciprocal Rank Fusion (RRF).

I

Intent Classification Determining what a user wants from their message. Agents use this to route to the correct specialist, tool, or workflow.

J

JSON Schema A machine-readable schema that defines required fields, types, and constraints for JSON data. Used by structured outputs to enforce response shape.

K

Kernel (Semantic Kernel) The central orchestration object in Semantic Kernel. Holds AI services, plugins, memory, and planners. Entry point for all SK operations.

L

Language Model API (VS Code) vscode.lm — an API that lets VS Code extensions use the LLM powering Copilot, without needing a separate API key. Available in VS Code 1.90+.

LLM Large Language Model — a deep learning model trained on large text datasets to understand and generate human language. Examples: GPT-4o, Claude 3.5, Llama 3.

M

MCP (Model Context Protocol) An open standard by Anthropic for connecting AI models to external tools and data sources via a structured protocol. See Lab 012 — What is MCP?.

MCP Server A process that exposes tools and resources to AI agents via the MCP protocol. Can be local or remote. Written in Python, TypeScript, C#, etc.

Memory (agent) The ability of an agent to store and recall information. Types: in-context (within the prompt), external (vector DB, key-value store), episodic (conversation history).

Multi-Agent System An architecture where multiple specialized agents collaborate to solve a task. Includes an orchestrator that routes requests and specialists that handle domain-specific work.

O

OBO (On-Behalf-Of Flow) An identity pattern where a service exchanges a user's token for another token to call downstream APIs with delegated user permissions.

Observability The ability to understand what an agent system is doing through traces, metrics, logs, evaluations, and cost signals.

Ollama An open-source tool for running LLMs locally on your laptop. Supports Llama, Mistral, Phi, Gemma, and 100+ models. Completely free, no internet required.

Orchestrator An agent whose primary role is to understand user intent and delegate sub-tasks to specialist agents or tools.

P

pgvector An open-source PostgreSQL extension that adds vector data types and similarity search. Enables storing embeddings directly in a Postgres database.

Planner (Semantic Kernel) A component that breaks a high-level goal into a sequence of steps, each mapped to an available plugin function. Examples: FunctionChoiceBehavior.Auto().

Plugin (Semantic Kernel) A class with methods decorated with @kernel_function. Exposed to the LLM as callable tools. Equivalent to "tools" or "functions" in OpenAI terminology.

Prompt Engineering The practice of crafting effective prompts to guide LLM behavior. Includes system prompts, few-shot examples, chain-of-thought, and output formatting. See Lab 005.

Prompt Injection An attack where malicious content in external data (documents, user input) overwrites the agent's instructions. A key security concern for RAG and agentic systems.

R

RAG (Retrieval Augmented Generation) A pattern that enhances LLM responses by first retrieving relevant documents from a knowledge base, then including them in the prompt. See Lab 006 — What is RAG?.

Reranking A second-pass step after initial retrieval that reorders results by relevance using a cross-encoder model. Improves precision at the cost of additional latency.

Row Level Security (RLS) A PostgreSQL feature that restricts which rows a database user can see. Policy defined in SQL: CREATE POLICY ... USING (customer_id = current_user). See Lab 032.

S

Semantic Kernel (SK) Microsoft's open-source SDK for building AI agents in Python, C#, and Java. Provides kernels, plugins, memory, planners, and multi-agent patterns.

Semantic Search Finding results based on meaning/intent rather than exact keyword match. Uses vector similarity (cosine, dot product) on embeddings.

Sparse Embedding A high-dimensional vector where most values are zero. Used in keyword/BM25 search. Contrasts with dense embedding.

Streaming Returning LLM output token by token as it's generated, rather than waiting for the full response. Improves perceived latency.

Structured Outputs Model outputs constrained to match a schema, usually JSON Schema or a Pydantic model, so downstream systems receive predictable fields and types.

System Prompt Instructions given to the LLM at the start of a conversation that define its persona, capabilities, and constraints. Not visible to the user in most UIs.

T

Temperature A parameter (0.0–2.0) controlling LLM randomness. 0 = deterministic (best for structured output), 1 = balanced, >1 = more creative/random.

Token The basic unit of text processed by an LLM. Roughly 3/4 of a word in English. Costs and context limits are measured in tokens.

Tool Calling See Function Calling.

Top-K / Top-P Sampling strategies for LLM output. Top-K: only consider the K most likely next tokens. Top-P (nucleus sampling): consider tokens whose cumulative probability exceeds P.

V

Vector Database A database optimized for storing and querying embeddings by similarity. Examples: pgvector, ChromaDB, Pinecone, Weaviate, Azure AI Search.

Vector Search Finding the nearest vectors (most similar embeddings) to a query vector using metrics like cosine similarity or L2 distance.

W

Webhook An HTTP endpoint that receives events from an external service. GitHub Copilot Extensions use webhooks to receive chat messages and return responses.

Common Acronyms

Acronym Full Name
AI Artificial Intelligence
API Application Programming Interface
BM25 Best Match 25 (keyword scoring algorithm)
CE Conformité Européenne (safety certification)
CLI Command Line Interface
CV Computer Vision
DI Dependency Injection
DSL Domain-Specific Language
GPT Generative Pre-trained Transformer
JSON JavaScript Object Notation
JWT JSON Web Token
LLM Large Language Model
MCP Model Context Protocol
ML Machine Learning
NLP Natural Language Processing
RAG Retrieval Augmented Generation
RLS Row Level Security
RRF Reciprocal Rank Fusion
SDK Software Development Kit
SK Semantic Kernel
SQL Structured Query Language

Missing a term? Open an issue or submit a PR!