Glossary

Plain-language definitions of the terms used across these docs.

Agent : A reusable configuration — a "role" — made up of instructions (a prompt), a chosen model, and the set of tools it can use. See Build & share agents.

Agentic dimension : The capacity for a product's features and actions to be driven by AI agents. The direction for our products is to gain this through Kiwi — see Powered by Kiwi.

Collection : A named group of documents you can attach to an agent, a session, or a single message to give it the right context. See Ground agents in your documents.

Connector (MCP) : A way to plug an external or project tool into Kiwi so it becomes available to any agent, built on the Model Context Protocol. See Extend with connectors.

Context window : The amount of text a model can take into account at once — the conversation, any attached documents, and the response all share this budget, measured in tokens. A larger context window lets an agent work with more material in a single turn.

MCP (Model Context Protocol) : An open standard for connecting AI applications to external tools and data sources. In Kiwi, connectors use MCP to make an outside tool available to agents. See Extend with connectors.

Media Studio : The part of the core that generates images, video, and audio from natural language. See Create media with Media Studio.

Memory : The conversation history Kiwi keeps for a session, so an agent remembers context across turns.

Model : The underlying AI model an agent uses to reason and respond. See Choose the right model.

Prompt : The instructions that define how an agent behaves.

RAG (retrieval-augmented generation) : Answering a question by first retrieving the relevant passages from a set of documents and supplying them to the model, so the response is grounded in those sources and can cite them. In Kiwi, this is how document grounding works — see Ground agents in your documents.

ReAct : The reason–act–observe loop an agent follows: it reasons about the task, acts by calling a tool, observes the result, and repeats until it can answer. Short for reasoning and acting. See Key concepts.

Reasoning : The model's internal "thinking" (chain-of-thought) before it answers. On supported models you can set the reasoning effort, trading speed for depth on harder tasks.

Session : A persistent conversation between a person and an agent.

Skill : A packaged set of instructions that teaches an agent how to perform a specific task, in a repeatable way, loaded only when needed. Distinct from a tool: a tool is what an agent can do; a skill is the know-how for getting a task done. Coming to Kiwi — see the Roadmap.

Streaming : Delivering a response live, piece by piece (token by token) as the agent works, rather than waiting for the whole answer.

Sub-agent : An agent that another agent delegates to, enabling more complex multi-agent workflows. On the Roadmap.

Token : The unit of text a model reads and produces — roughly a short word or word-piece. A model's context window is measured in tokens.

Tool : A capability an agent can call to do something concrete — web_search, spreadsheet_query_sql, generate_image, code_interpreter, and so on. Tools are what an agent can do. See the full list in Features.