AI Playground

Open /ai/playground (or /ai with the playground tab) to test a request before putting it into a work process. It is available only when the user may run AI; when canRunRequests is false, the tab is not shown.

What you can test

• a direct completion or a selected scenario;
• text or JSON output;
• template, messages, or hybrid input mode;
• an optional connection or the module default;
• JSON context and a streaming response;
• anonymized improve-text, short-summary, and JSON-analysis presets.

For message mode, add at least one user or assistant message. Fix empty fields, invalid JSON, and unknown scenarios before running.

How the request is built

In template mode, fill in the prompt and optional context. In message mode, add a non-empty list of user or assistant messages. Hybrid mode combines prompt, context, and messages. A scenario uses the fields for its selected mode; streaming is available for scenarios. Check every visible form field before running.

/ai/playground has no @ picker: it accepts manually entered prompts, messages, and JSON context. Test mention context in the direct Assistant chat at /chat/:conversationId: use only synthetic Anna Test Employee and CRM opportunity Rebranding — demo, choose the person from the @ suggestion instead of typing a name, and send a read-only request. Until the UI fix lands, a selected mention in the direct Assistant branch may be stored as plain text without its ID; @Display Name in recent context is therefore not proven. Reject raw IDs, {}, mock_default, mock-v1, and raw JSON, and real entities; public screenshots or behavior claims require a frontend regression, deployment, and readable display-name proof with no mutation.

Limits and overflow

Check the visible input, output, and context limits, timeout, result format, and overflow behavior. With “Reject,” an oversized request does not run; with “Truncate,” the interface should clearly show that part of the context was shortened.

For a streaming response, distinguish the beginning, intermediate text, completion, stop, and error. Intermediate text is not a confirmed result until completion. If an error shows technical server or transport text instead of a localized explanation, do not copy or publish it: it is not a safe reason for a decision or evidence. Record the route and time, then ask the portal owner without the error contents; do not blindly run the request again.

Read the result panel

The right panel shows status, latency, selected connection and model, input mode, and any truncation warning. Check the text or JSON result; use a streaming response only after it has completed.

Access and permissions

Permission to run AI does not grant permission to change connections or scenarios. Without run access, the Playground tab is hidden; do not work around the restriction.

Safety

Use synthetic names, amounts, and records. Do not paste keys, passwords, payment or personal data, or private files. The Playground prepares a draft; verify financial values in the source system.

Before reusing a result

1. Confirm the scenario and connection.
2. Check limits, format, and truncation.
3. Verify facts, names, deadlines, and promises against the source context.
4. Put the reviewed result into a task, CRM, or document; an AI response is not a decision by itself.

Related: AI Hub, AI scenarios, AI connections.