Glossary
Introduction
This glossary provides brief definitions of key agentic and generative AI terms used in the Mendix documentation. To explore working implementations of these concepts, see the GenAI Showcase App, which demonstrates RAG, tool calling, MCP integration, and more.
Agent
An AI system that combines an LLM with a reasoning loop and can be extended with tools. It can plan, evaluate results, and iterate toward a goal.
In Mendix, you can build agents using Agents Kit, a collection of starter apps, connectors, and modules that support implementations from simple text generation to complex multi-step agentic workflows. For more information, see GenAI Agents.
Agent Orchestration
The coordination layer—often built within an app or workflow—that manages how agents work together: which agent takes responsibility, what context is shared, and how handoffs, retries, failures, and recovery are handled.
Agentic
Describes a system, workflow, or behavior that operates with some degree of intelligence, autonomy, or adaptive decision-making, even when not built as a standalone agent.
Agentic Workflow
A workflow pattern where specific steps are handled through agentic decision-making instead of fixed logic. The workflow keeps the overall structure and control; agentic components handle steps that require reasoning, interpretation, or adaptive action.
Embeddings
Mathematical representations of text as numeric vectors, where conceptually similar text strings result in similar vectors. This enables semantic search based on meaning rather than exact character matches. Embeddings are typically stored in a vector database and are a core component of RAG.
Generative AI (GenAI)
A category of artificial intelligence technology that can generate new content such as text, images, code, or other outputs based on learned patterns from training data. Generative AI can be used to accelerate business processes, provide user-friendly interactions, and enhance apps with features such as chatbots, content generation, text analysis, image analysis, and language translation.
Hallucination
When an LLM generates incorrect, nonsensical, or fabricated information that is not grounded in its training data or the provided context. For example, an LLM might confidently state that a fictional person won an award that never existed, or cite nonexistent research papers. To reduce hallucinations, techniques like RAG ground responses in verified data, and patterns like human in the loop add validation steps for critical decisions.
Human in the Loop
A design pattern where an agent pauses at defined points to request human input such as approval, correction, or a decision. The human actively participates in the agent’s execution rather than passively observing.
In Mendix, this is implemented through user access approval settings on tools, which control when tools get executed and whether they are visible to the user. For example, a tool that deletes records might require explicit user confirmation before execution. The ConversationalUI module supports human-in-the-loop interactions out of the box. For implementation details, see Human in the Loop.
Knowledge Base
A storage system for discrete pieces of information that can be retrieved and used to augment prompts. Knowledge bases are essential for RAG patterns. When knowledge is stored as embeddings in a vector database, semantic similarity searches are possible.
Mendix supports managed cloud knowledge base services such as Mendix Cloud GenAI Knowledge Base Resource Packs and self-managed implementations using the PgVector Knowledge Base module.
Large Language Model (LLM)
A neural network trained on large amounts of text that can understand input and generate responses. LLMs can perform tasks such as text generation, answering questions, summarization, and translation.
Mendix connectors support LLMs from providers including OpenAI, Amazon Bedrock, and Mendix Cloud GenAI. For details on supported models, see Available Models.
Model Context Protocol (MCP)
An open protocol that standardizes how large language models can connect to applications. MCP enables apps to expose tools and prompts to external clients, and consume tools and prompts from external servers.
Mendix provides an MCP Server module to build an MCP server from a Mendix app, enabling developers to expose tools and prompts to external MCP clients. Mendix also provides an MCP Client module that enables your app to connect to MCP servers and discover and use their tools and prompts. For more information, see Model Context Protocol.
Prompt Engineering
The practice of structuring instructions to guide generative AI models to produce desired outcomes. Prompt engineering involves designing the input text sent to the LLM, typically through system prompts and user prompts. This includes crafting instructions on what the model should do, providing context and information needed to follow those instructions, structuring input data, and specifying the requested output format. The quality of prompts directly influences the quality of AI responses. For more information, see Prompt Engineering.
Retrieval Augmented Generation (RAG)
A pattern that combines knowledge retrieval with text generation. RAG retrieves relevant information from a knowledge base, augments the prompt with that information, and generates a response grounded in the retrieved data. This allows models to answer questions about private or domain-specific data. For example, when a user asks "What is our refund policy?", RAG retrieves relevant policy documents from the knowledge base, adds them to the prompt, and generates an accurate answer based on the actual policies.
System Prompt
Instructions that define the model's behavior, role, and constraints. System prompts establish how the LLM should respond and what personality or expertise it should adopt. These are typically set by developers and stay the same across interactions. For example, "You are a helpful customer service assistant for an e-commerce platform. Always be polite, concise, and follow company policies when answering questions."
Token
The basic unit of data that LLMs process. For text input, models break down text into tokens, which can be words, parts of words, or individual characters depending on the model's tokenization method. For example, "agentic workflow" might be broken into tokens like ["agent", "ic", " work", "flow"].
Tool Calling
Also known as tool use or function calling, a capability that allows LLMs to call specific tools to gather additional information or perform actions. The model returns a structured request indicating which tool to call and with what parameters, which the app then executes and returns results to the model. Tool calling enables agent patterns like ReAct (Reasoning + Acting), where agents iteratively reason about what to do and then call tools to take action. For more information, see Tool Calling.
Trace
A structured record of execution across a system or workflow. In the context of an agent, a trace captures the sequence of inputs, reasoning steps, tool calls, intermediate actions, and outputs that led to a result. It also records token consumption, the duration of each step, and whether the execution was successful.
In Mendix, tracing is supported by the GenAI Commons module to help debug and monitor agent behavior.
User Prompt
The input text provided by an end-user or app to an LLM. User prompts contain the specific question, request, or data that the model should process. These vary with each interaction based on what the user needs. For example, "What is your refund policy?" or "Summarize this customer feedback." User prompts work in combination with system prompts to produce responses.
Vector Database
A specialized database designed to store and query embeddings. Vector databases support similarity searches using mathematical techniques to find the most semantically similar items to a query. This capability is essential for RAG patterns.
In Mendix, PostgreSQL with the pgvector extension serves as a vector database. For more information, see PgVector Knowledge Base.