Skip to main contentAgent
An Agent is a queryable AI application that retrieves relevant data from connected datastores and generates responses grounded in that data. Agents are built in the Contextual AI app using Agent Composer: you choose a template (e.g. Basic Search, Agentic Search) or design a custom workflow with nodes and steps. Agents can do multi-step retrieval, use tools (search, APIs, MCP), and produce cited answers. You create and manage agents via the API or the platform UI.
Agent Composer
Agent Composer is the framework for building agents on Contextual AI. You compose workflows from building blocks (search steps, generation steps, tools, conditional logic) either by editing YAML or using the visual workflow builder. Self-serve users can use Basic Search and Agentic Search templates; enterprise users get access to the full builder and additional templates. See Agent Composer overview and Agent Composer templates.
Attributions / Citations
Attributions (citations) are in-line references that credit the sources used to generate a response. When you query an agent, attributions appear for factual claims (e.g. [1], [2]) and link to the retrieved documents. You can view them in the API response or in the UI. Agents built with Agent Composer can be configured to enforce citation format (e.g. number with a References section). See core system prompt and disable commentary.
Chunking
Chunking is the process of splitting documents into smaller, semantically meaningful pieces so the agent can retrieve and reason over them. Contextual AI supports multiple chunking modes to suit different document types. See Chunking configurations.
Connector
A connector links an external data source (e.g. Confluence, Google Drive, SharePoint, Box) to Contextual AI. Content is synced into one or more datastores and kept up to date. See Connectors overview.
Datastore
A datastore is a repository of ingested content (documents, chunks) that an agent searches when answering queries. You connect datastores to an agent when building the agent; one agent can use multiple datastores, and one datastore can serve multiple agents. Content is added via connectors, local upload, or the documents API. See Datastores reference.
Document
A document is a single unit of content (e.g. PDF, HTML, Markdown) ingested into a datastore. After ingestion, documents are parsed, chunked, and indexed for retrieval. You ingest documents via the API or via connectors.
Knowledge (retrievals)
Knowledge (or retrievals) is the set of chunks and sources the agent retrieved from its datastores to generate a response. In Agent Composer workflows, search steps output retrievals that are passed to generation steps. Responses can include attributions pointing to these sources.
LMUnit
LMUnit is an evaluation method that uses natural-language unit tests to check specific qualities of an agent’s response (e.g. “Is the response succinct?” or “Is the complexity appropriate for the audience?”). You define and run these tests via the LMUnit API. See also LMUnit overview.
Node
In Agent Composer, a node is a single step in the workflow graph (e.g. a search step, a generation step, a tool call). Nodes are connected by input mapping from outputs of one node to inputs of another. See YAML format and node reference.
Query / Prompt
The query (or prompt) is the user’s question or instruction sent to an agent. You submit queries via the query API or through the platform UI. The agent uses this input to drive retrieval and generation.
RAG
Retrieval Augmented Generation (RAG) is a technique that grounds LLM outputs in retrieved knowledge. Contextual AI agents use RAG: they search your datastores, then generate responses from the retrieved chunks, with citations. Agent Composer lets you configure retrieval (e.g. search steps, reranking, filters) and generation (e.g. system prompts, model choice) in one workflow.
Response
The response is the agent’s answer to a query. It is generated from retrievals and can include attributions and a References section. Response behavior is controlled by the agent’s workflow and generation config (e.g. max tokens, temperature).
System Prompt
The system prompt (and, when no retrieval is used, the no-retrieval system prompt) defines the agent’s behavior and how it should use retrieved content. You set these when building an agent in Agent Composer. See Core system prompt and No-retrieval system prompt.
Template
A template is a pre-built Agent Composer workflow. Basic Search and Agentic Search are available for self-serve users; additional templates (e.g. Device Log Analysis, domain-specific templates) are available for enterprise. See Agent Composer templates and Template catalog.
Workspace
A workspace is the top-level container for your agents, datastores, API keys, and other resources. Access and billing are managed at the workspace level. 
