AI Assistant

What is the AI Assistant?

The AI Assistant is a dockable panel with provider configuration, a chat log and a prompt box. Generated plans can be reviewed and inserted with one click.

The AI Assistant is a dockable panel in the desktop application that connects the designer to a large language model – a local Ollama instance or OpenAI / ChatGPT (and any OpenAI compatible endpoint). You type a requirement in natural language, e.g. “Read all CSV files from the inbox folder, send each one as an e-mail attachment and move it to an archive folder afterwards”, and the assistant answers with a workflow plan: a set of tasks with their properties and the connections between them.

The plan is not an opaque text answer. It is a strictly structured proposal that the designer validates against the tasks currently available in the toolbox – built-in tasks, tasks from referenced NuGet packages and your own custom tasks alike – and then materialises as real nodes on the canvas. Nothing is executed and nothing is saved until you decide to.

How It Works

1. When you send a prompt, the designer builds a task catalog from the toolbox: full type name, display name, description, category, outputs (e.g. Then / Else, Iteration / Conclusion), input/output definitions and every editable property with its type and allowed values.
2. Optionally the current workflow is summarised as well – existing nodes, their identifiers, free outputs and connections – so the model can append to an existing branch instead of starting from scratch.
3. Catalog, workflow summary and a fixed set of composition rules (expression syntax, scope semantics, single input per task, how to merge branches) are sent as the system prompt together with your chat history.
4. The model has to reply with a single JSON plan. Markdown fences or surrounding prose are tolerated and stripped.
5. The plan applier resolves every task type against the catalog, creates the task instances through the regular node factory, assigns the proposed property values (with type conversion for enums, numbers, booleans and TimeSpans), wires the connections and lays the new nodes out left-to-right.
6. Problems – unknown task types, unknown properties, invalid connections – are reported as warnings under the chat entry and fed back into the conversation so the next request can correct them.

A generated plan after insertion: the tasks appear as regular nodes, grouped and connected, ready to be fine-tuned in the Properties panel.

Opening and Configuring the Panel

Open the panel via Panels → AI Assistant in the ribbon. The Configuration expander at the top is opened automatically until a provider is fully configured.

Press Save to persist the configuration. It is stored per user in %LOCALAPPDATA%\ultimate networX\AURORA Workflows\ai\assistant.json.

Working with the Chat

• Enter sends the prompt, Shift+Enter inserts a line break. Cancel aborts a running request, Clear starts a new conversation.
• Each assistant reply lists the proposed tasks with their key properties and the connections as a readable summary. Replies that contain tasks show an Insert into workflow button and the task count.
• After insertion the entry shows how many tasks and connections were created plus any warnings. The conversation keeps its history, so you can iterate: “use a Decision instead of the Loop”, “add error handling around the mail task”, “rename the variable to invoiceFiles”.
• If the model needs more information it can ask a clarifying question instead of returning tasks; simply answer in the prompt box.
• Select a node before sending a prompt to tell the assistant where the new flow should be attached. Without a selection new flows are connected to the Start node.

The plan format understood by the applier looks like this:

{
  "summary": "Reads CSV files from the inbox, mails each one and deletes it afterwards.",
  "entry": "t1",
  "tasks": [
    { "id": "t1", "type": "Aurora.Workflows.Tasks.GetFileListTask", "name": "Find CSV files",
      "outputVariable": "csvFiles",
      "properties": { "DirectoryPath": "C:\\Inbox", "Pattern": "*.csv", "SearchOption": "TopDirectoryOnly" } },
    { "id": "t2", "type": "Aurora.Workflows.Tasks.ForEachTask", "name": "For each file",
      "inputVariable": "csvFiles",
      "properties": { "ElementVariableName": "file" } },
    { "id": "t3", "type": "Aurora.Workflows.Tasks.SendMailTask", "name": "Send file",
      "properties": { "Host": "smtp.example.com", "Sender": "robot@example.com", "Recipient": "accounting@example.com",
                      "Subject": "=\"CSV import \" + Path.GetFileName(file)", "Body": "=File.ReadAllText(file)" } },
    { "id": "t4", "type": "Aurora.Workflows.Tasks.DeleteFileTask", "name": "Delete file",
      "properties": { "FilePath": "=file" } }
  ],
  "connections": [
    { "from": "t1", "to": "t2" },
    { "from": "t2", "output": "Iteration", "to": "t3" },
    { "from": "t3", "to": "t4" }
  ]
}

Property values follow the normal designer semantics: plain text is used literally, a value starting with = is a C# expression evaluated at runtime, and variables declared via outputVariable can be referenced by name inside expressions – exactly like in the Data Context and the Properties panel.

Composition Rules the Assistant Follows

Privacy and Cost

• With Ollama everything stays on your machine or in your network. No data is sent to a third party.
• With OpenAI the task catalog, the workflow summary (when enabled) and your prompts are sent to the configured endpoint. Property values of existing nodes are included in the summary – disable Send current workflow as context if the open workflow contains sensitive literals.
• The task catalog can be sizeable when many packages are loaded; every request includes it in full. Use Max. output tokens and a smaller model for quick iterations, a stronger model for the final composition.
• The API key is protected with Windows DPAPI for the current user account. Copying assistant.json to another machine or user will not expose it – it simply has to be entered again.

Troubleshooting