AI terms - Voiceflow Documentation

Term	What it is
Model	The language model generating a reply. Which one an agent uses is a behaviour setting, and the choice affects both quality and credit cost.
LLM	Large language model. Used interchangeably with model above.
Token	The unit a model reads and writes text in, roughly a word-piece. Usage and cost are measured in tokens, which is why analytics reports them.
Context window	How much text a model can consider at once, counted in tokens. Everything the agent sends on a turn competes for the same window.
Prompt	The text given to a model to produce a reply. In Voiceflow this is assembled from the global prompt, the instructions, and the conversation so far.
System prompt	The instruction layer a model treats as standing orders rather than user input. Voiceflow’s global prompt fills this role.
RAG	Retrieval-augmented generation. Fetching relevant documents and giving them to the model so the answer is grounded in your content rather than the model’s training. This is what the knowledge base does.
Embedding	A numeric representation of a piece of text, used to find passages that mean something similar rather than merely sharing words.
Chunk	One passage a document is split into before it is embedded. Chunking decides what retrieval can return.
Grounding	Constraining an answer to supplied source material, so the model reports rather than invents.
Hallucination	A fluent, confident answer that is not true. The reason grounding and evaluations matter.
Tool calling	A model choosing to invoke a tool, with arguments, instead of replying in prose. What makes a playbook agentic.
Temperature	How much randomness a model uses when choosing words. Lower is more repeatable, higher more varied.
MCP	Model Context Protocol. An open standard for exposing tools to a model, so the same tool works across clients.

Term

What it is

Model

The language model generating a reply. Which one an agent uses is a behaviour setting, and the choice affects both quality and credit cost.

LLM

Large language model. Used interchangeably with model above.

Token

The unit a model reads and writes text in, roughly a word-piece. Usage and cost are measured in tokens, which is why analytics reports them.

Context window

How much text a model can consider at once, counted in tokens. Everything the agent sends on a turn competes for the same window.

Prompt

The text given to a model to produce a reply. In Voiceflow this is assembled from the global prompt, the instructions, and the conversation so far.

System prompt

The instruction layer a model treats as standing orders rather than user input. Voiceflow’s global prompt fills this role.

RAG

Retrieval-augmented generation. Fetching relevant documents and giving them to the model so the answer is grounded in your content rather than the model’s training. This is what the knowledge base does.

Embedding

A numeric representation of a piece of text, used to find passages that mean something similar rather than merely sharing words.

Chunk

One passage a document is split into before it is embedded. Chunking decides what retrieval can return.

Grounding

Constraining an answer to supplied source material, so the model reports rather than invents.

Hallucination

A fluent, confident answer that is not true. The reason grounding and evaluations matter.

Tool calling

A model choosing to invoke a tool, with arguments, instead of replying in prose. What makes a playbook agentic.

Temperature

How much randomness a model uses when choosing words. Lower is more repeatable, higher more varied.

MCP

Model Context Protocol. An open standard for exposing tools to a model, so the same tool works across clients.

Capitalisation carries meaning here. “Agent step” and “Function step” name specific steps; the bare words step and tool stay lowercase.