Dashboard Workflow Editor

The dashboard (server/packages/dashboard) is where operators configure orchestration without redeploying code.

Tabs at a glance

• Projects & Devices: assign a workflow to each device.
• Workflow Library: create/edit graph workflows.
• Documents: browse uploads and run embedding-based search.

Building workflows

Open Workflow Library and create a new workflow or edit an existing one.

Node types

Node type	Purpose	Required config
worker	ReAct-style worker agent	Model, optional system message, optional tools
supervisor	Orchestrates worker members	Model, config.members (worker node keys)
tool	Single tool invocation node	Exactly one tool, optional IO mapping

Assigning tools to workers

1. Select a worker node.
2. In Assigned tools, toggle one or more tools.
3. Save the workflow.

Workers can hold multiple tool assignments.

Tool node mappings

For tool nodes, map tool schema fields to graph state keys:

• input_mapping: graph state -> tool input
• output_mapping: tool output -> graph state

Literal input values can be passed with _const: (for example _const:image/png).

Validation rules before save

The canvas enforces:

• Unique node keys
• Valid node types (worker, supervisor, tool)
• Model required for worker/supervisor
• Exactly one tool on tool nodes
• Valid supervisor membership (workers only, no duplicates)
• At least one edge to END
• Entry node must have a path to END

If save fails, fix the message shown in the editor and retry.

Assign workflow to devices

After saving a workflow:

1. Go to Projects & Devices.
2. Select a workflow for each device.
3. Click Apply to persist assignment.

This controls which graph runs for that device’s uploaded documents.

Semantic document search

In Documents, search with natural language (example: red bike by a window).

• If the API finds a matching description seed, results are ranked by embedding distance.
• If no seed is found, it falls back to lexical matching.
• Matched cards show a similarity badge (for example 92% match).

Use this view to validate that the ingestion + embedding pipeline is producing useful retrieval behavior.