Skip to main content
Chats and agents run on the same grounded-answer runtime. The runtime combines conversation context, knowledge-base retrieval, connected data, memory, and tools, then cites supporting sources whenever retrieval finds them.

Quick and Deep modes

Quick

A fast, single-pass path for direct questions. It retrieves relevant knowledge-base content and returns an answer with minimal orchestration.

Deep

A multi-step agent loop for questions that need repeated searches, connected data, charts, apps, tools, or delegation.
Saved agents always run in Deep mode. Chats can select quick or deep with the agent_mode request field.

The Deep agent loop

1

Plan

The model interprets the message, available context, and tool catalog, then chooses the next useful action.
2

Call a tool

It can search the knowledge base, run read-only SQL, build a chart, call a connected app or custom tool, use an MCP server, or delegate to another agent.
3

Observe

Tool results return to the model as grounded context. The model can refine its plan and call another tool when needed.
4

Answer

The loop ends with a final answer and cites sources whenever retrieval produced supporting passages. A configurable step limit keeps the run bounded.

Runtime capabilities

CapabilityQuickDeep
Knowledge-base retrievalSingle-passRepeated, agent-directed searches
Source citationsWhen retrieval finds supporting passagesWhen retrieval finds supporting passages
Read-only connected database queriesYes
Charts from connected dataYes
Connected apps and custom toolsYes
MCP toolsYes
Agent delegationYes
The exact tools available in a Deep run depend on the knowledge base, chat, or agent configuration and the caller’s access. Connected apps must be authenticated before the runtime can use them.

Streaming a run

Streaming endpoints expose the run as it happens. Events can include progress, tool activity, final-response content, sources, charts, token usage, completion, and errors. The main runtime normally emits one canonical final_response with isComplete: true; clients should also support partial final_response events and accumulate their content for protocol compatibility. The non-streaming endpoints use the same runtime and return after the final result is ready. See Chat lifecycle and Chat & streaming for request and event details.

Grounding and context

The runtime does not place the entire document corpus into the prompt. Instead, it receives a short preview of available content, searches at run time, and adds the retrieved passages as tool results. Conversation summaries and durable memory are separate context layers. Read Prompt assembly for precedence and RAG lifecycle for retrieval behavior.

Next steps

Prompt assembly

Understand prompt layers, memory, history, and custom overrides.

RAG lifecycle

Follow a knowledge-base search from query to cited source.