Agent builder
Build quickly in simple mode, or design a custom agent with the visual node editor in advanced mode. Validate it in preview, then publish it for use in chat and embeds.
The agent builder is where you design a RAG chatbot. You configure the answering LLM, document retrieval tools, and the conversation flow all in one place. There are two ways to start: simple mode (recommended), which builds a working agent from a short form, and advanced mode, which lets you wire nodes together to design the flow yourself. Whichever you choose, you can keep editing the agent in the builder afterward.
Create new agents from the Agent builder menu in the sidebar, and use the list to edit, run, or delete existing ones.
Simple mode
Just pick a name, a model, and the document groups to search. You get a ready-to-use agent without needing to know anything about nodes.
Advanced mode
Drag and connect 28 nodes across 8 categories. Design complex flows with branching, parallel steps, and transforms yourself.
Preview chat
Test conversations right inside the builder before publishing. Inspect the steps that ran and the sources used via the debug timeline.
Publish and share
Once published, the agent appears in chat. Set the share scope to the whole organization or specific teams, and optionally enable embedding.
Which mode should you start with
When you create a new agent, a mode picker appears first. Simple mode is enough for most cases; choose advanced mode when you need finer control. The mode is only a starting point: every agent is edited in the same builder afterward.
| Aspect | Simple mode | Advanced mode |
|---|---|---|
How you start | Generated from a form | Place nodes on a blank canvas |
Best for | First-timers, standard RAG chatbots | Power users needing branching or tool mixes |
What you configure | Model, retrieval tools, document groups | All 28 nodes across 8 categories |
Editing later | Add and edit nodes in the builder | Add and edit nodes in the builder |
If this is your first agent, start with simple mode. Agents created in simple mode can still be extended freely by adding nodes in the advanced builder later.
Build an agent in simple mode
Simple mode turns a single form into a working agent. Based on your inputs, it automatically connects the input, model, memory, agent, and output nodes along with the retrieval tools you enabled.
Enter a name and description
Give the agent a name and, optionally, a description. The name identifies the agent in the chat list and in the builder.
Choose an LLM model
Pick the LLM provider and model that will generate answers. Available models are based on the API keys registered in your organization settings. If you do not see a model, register the provider in organization settings first.
Turn on retrieval tools
You can independently enable document (text) search, image search, and knowledge graph search. Turning on any of them prompts you to select the document groups to search.
Set reranking and share scope
When text search is on, you can also configure reranking to improve result accuracy. The default reranker is Qwen3-Reranker, and Cohere is available if your organization has a Cohere key registered. Finally, decide whether to share this agent with the whole organization or only specific teams.
Create and open the builder
By default the agent is published immediately (you can turn this off in the create dialog); a published agent appears in chat right away, while a draft is only testable in the builder preview. After creating, you go straight to the builder, where you can test in preview, add more nodes, or change the publish state.
With no retrieval tools enabled, the agent behaves like a plain LLM chatbot. To make it answer from your uploaded documents, enable at least document search and select a document group.
Advanced mode: the visual node editor
Advanced mode lets you design the conversation flow by wiring nodes on a blank canvas. There are 28 nodes across 8 categories: drag them from the palette onto the canvas and connect their ports with edges. Click any node to adjust its settings in the right-hand panel.
| Category | Nodes | Purpose |
|---|---|---|
Input | 3 | Chat input, webhook input, and text input as flow entry points |
Output | 3 | Send results out via chat output, data output, or print |
Models | 1 | The LLM model node sets the provider, model, and parameters |
Agent | 3 | AI agent, deep agent, and subagent perform the reasoning |
Memory | 1 | Simple memory that keeps conversation context |
Tool | 8 | Document, image, and graph search, guardrails, function, API, MCP, and web search |
Flow | 5 | Condition, router, parallel, merge, and classify control the flow |
Transform | 4 | Template, JSON parser, type convert, and filter shape the data |
On the canvas, port colors indicate data types, and incompatible ports cannot be connected. Connections that would create a cycle are also blocked, keeping the flow safe to run. Your edits are tracked automatically so you can undo and redo.
- UndoCtrlZ
- RedoCtrlShiftZ
Retrieval and tool nodes only take effect when connected to the agent node's tools port. A node left unconnected is ignored at run time, so always confirm the behavior in preview before publishing.
Preview and publish
Switch between edit and preview modes at the top of the builder. In preview mode you chat with the agent exactly as users would, validating its behavior. When you are satisfied, publish it to expose it in chat.
Save
Your edits show a save status automatically. When there are unsaved changes, save them from the builder header, and confirm the agent is saved before publishing.
Test in preview
Switch to preview mode and ask a few questions. Alongside the answer, the debug timeline and source cards show which nodes ran and which documents were cited. You can also attach files to check multimodal behavior.
Publish
Press the publish button in the builder header to mark the agent as published. Only published agents appear in the chat agent list where users can start conversations. Unpublishing returns it to draft status.
A published agent is ready to use in chat, shows feedback statistics, and unlocks embed settings. While in draft, it can only be tested through the builder preview.
Sharing and permissions
When you create an agent, you set its share scope to the whole organization or to specific teams. Sharing with the whole organization lets every member see and use it; sharing with a team limits it to that team's members. The share scope can be changed later.
| Permission | Viewer | Member | Admin | Owner |
|---|---|---|---|---|
| View and chat with shared agents | ||||
| Create new agents | ||||
| Edit in the builder | ||||
| Publish and unpublish | ||||
| Delete agents |
Organization and team boundaries are enforced on the server. Even when published, an agent without embedding enabled cannot be reached externally, and users only see agents from the organizations and teams they belong to.
Frequently asked questions
Yes. The mode is only a starting point. An agent built in simple mode can be extended into an advanced flow by adding and editing nodes in the builder. No conversion step is needed.
The model list is populated from the AI provider API keys registered in your organization settings. If a model is missing, register that provider and key in organization settings first.
Check that the document search tool is on and the correct document group is selected. The documents also need to have finished indexing to be searchable. If the source cards are empty in preview, review the retrieval tool connection and group settings.
It follows the share scope. If shared with the whole organization, every member can use it in chat; if shared with a team, that team's members can. To use it on an external website, you also need to enable embedding separately.