2. AI Analysis playground with Prompt

2.3. Custom Prompt: User can define a pre-prompt. After that, click Analysis Source to work with Custom Prompt

3. How to analyze content 

A new unified AI Workbench

The AI Workbench is a unified workspace that brings all of EvolveIT's AI capabilities together in one place. From a single asset-focused tab you can pick the language model, run built-in or custom prompts, view and edit source code, hold a streaming chat with the AI, and keep a searchable, replayable history of every conversation, including the tools the AI used along the way.

This document walks through opening the Workbench and using each of its main capabilities. The examples assume you are logged in to the EvolveIT Dashboard with an inventory loaded.

1. Multi-tab workspace with AI Model selection

The AI Workbench opens as its own window containing a tab strip. Each tab is a self-contained session focused on one inventory object, with its own source code, chat, settings, and history. You can keep several tabs open at once, for example one per program you are analyzing, and switch between them without losing context.

The Workbench window can be minimized (it collapses to the AI Workbench button in the toolbar, which shows a badge with the number of open tabs) and reopened without losing your tabs. Closing a tab ends that session; closing the last tab closes the window.

1.1 Opening the AI Workbench

There are several ways to open the Workbench, depending on where you are working.

From the left menu

Go to Dashboard → AI Feature → AI Workbench. This opens the Workbench so you can start a session and paste or type the content you want to work with.

From an object (asset) browser

In any object browser, use the AI action on an object's row. A Workbench tab opens already focused on that object. Its name, type, RID, and description are filled in automatically, and its source code is loaded for you.

Figure 2: Opening the AI Workbench for an object from the browser.

From the source code viewer

The source code viewer can be reached from an asset browser, a graph/diagram view, or a text search result. With the code open, you can optionally select a region of code, then right-click in the editor and choose AI Workbench from the context menu. A Workbench tab opens focused on that program with its source ready to analyze.

From an Asset Group

Opening the Workbench on an Asset Group shows the group's members in a Sub Assets list (see section 3) so you can pick which members to focus on; opening it on an individual sub-asset shows that asset's source code.

1.2 Re-opening the same object

If you open the Workbench for an object that already has a tab, the Workbench simply switches to the existing tab instead of opening a duplicate, so each object has at most one live session.

1.3 Selecting the AI Model

At the top of each tab, the Request Information panel holds the controls that govern the request:

• AI Model: a drop-down listing every available language model by name. Pick the model you want this session to use (see section 9).
• Mode: switches between Built-in Prompts and Custom Prompts (see section 2).
• Use Memory: when checked, the AI uses the running conversation as context so follow-up questions are understood in relation to earlier ones. Uncheck it to treat each request on its own.

The Asset Information panel beside it shows the object the tab is focused on (RID, Type, Name, Description). This context is supplied to the AI with every request.

Figure 4: The Asset Information and Request Information panels.

2. Custom Prompts mode

The Mode drop-down controls how you drive the AI. There are two modes.

2.1 Built-in Prompts + Chat

EvolveIT ships a curated library of analysis prompts. Choose a Category and then a Request Type from that category, and click Analyze Source to run the selected prompt against the current object. You can also type a free-form question in the chat box at any time.

2.2 Custom Prompts + Chat

Custom Prompts mode lets you drive the AI with your own pre-prompts instead of the built-in library. (This mode is available to users who have the Custom Prompt Management permission.) Select a saved prompt from the drop-down to preview it, then click Analyze Source to run it. Your custom instructions are sent ahead of your question so they steer how the AI responds.

To build and manage your prompts, click Manage Prompts to open the Advanced Prompt Editor:

• Add Prompt: add another prompt row and type its text.
• Remove: delete a prompt row.
• Save to File: export your prompt set to a JSON file for reuse.
• Load: import a prompt set from a previously saved JSON file.
• Apply: load the edited set back into the session's prompt drop-down.

3. Source code panel with "Expanded code" and "Edit Src Code" options

For source-bearing objects, the Source Code panel shows the object's code in an editor beside the chat session. The panel is collapsible, so you can widen the chat when you do not need to see the code. For an Asset Group, this panel is replaced by a Sub Assets list where you can tick the members you want the AI to focus on.

Two checkboxes control how the code is used in your requests:

• Edit Src Code: the source is shown read-only by default. Tick this box to make the editor editable; the code you have in the editor is then sent with your request, so you can analyze a modified or in-progress version rather than what is stored.
• Expanded code: available for application programs that can include copybooks. When ticked, the AI is told to use the expanded source (with copybook includes resolved); when unticked, it uses the original source. The option is hidden for object types that have no copybook includes.

4. Conversation history with accurate replay and tool-call support

A collapsible Conversation History sidebar runs down the left side of the tab. It lists earlier conversations for the current object, each showing a preview of the first question and the date.

Use the toolbar at the top of the sidebar to:

• Switch between the Mine and Shared tabs (see section 5).
• Start a fresh conversation with New.
• Reload the list with Refresh.

Click any entry to reopen it. Reopening a conversation replays it in order rather than dumping everything at once: each turn shows your question, then the tool calls the AI made (each transitioning from Running to Completed), and finally the AI's answer. This faithfully reproduces what happened in the original session, including which tools were used and what they returned.

Conversations are saved automatically when a response completes, so your history builds up without any manual step.

5. Save conversation settings; public/shared conversation support

5.1 Per-conversation settings

The Workbench remembers each conversation's settings (the selected AI Model, Mode, and Use Memory) and restores them when you reopen that conversation, so you do not have to reconfigure each time.

5.2 Sharing a conversation

Conversations can be shared with other users in the inventory. On a conversation you own, click the Advanced Options (gear) action and choose its visibility:

• Private: only you can see it.
• Public (Read Only): all users can view it, but no one else can add queries.
• Public (Read & Add Query): all users can view it and append their own queries.

The history sidebar's Shared tab lists public conversations along with the owner's name. When you open a conversation shared by someone else, a banner identifies the owner; in read-only shared conversations the input box is disabled so you can review but not modify them.

6. Markdown rendering for AI responses

AI responses are rendered as formatted Markdown rather than plain text. Headings, bold and italic text, lists, tables, block quotes, and links are all displayed with proper styling, and links open in a new tab. Code is shown in syntax-styled code blocks with a one-click Copy button, making it easy to lift generated snippets straight into your editor.

7. Stop button to interrupt long-running API requests

While a request is running, a status bar appears showing that the request is in progress, together with a Stop API button. Clicking it immediately aborts the request, marks any tool calls still in progress as Stopped, and returns the input box to you, so you never have to wait out a long or runaway request.

8. Download Session

Each conversation in the history sidebar has a Download Session action that packages the session as a ZIP file and downloads it to your machine, giving you a complete, portable record of the run for sharing, auditing, or troubleshooting. The ZIP contains:

• All output files the session produced, organized into folders by category: Trace Output (the field/code trace artifacts) and Modernization Output.
• A generated session overview page (session_overview.html) at the root of the ZIP. Open it in a browser for a summary of the session: the inventory and session identifiers, the total file count, and every output file grouped by category for quick navigation.

9. Set default LLM and select LLM by name

The AI Workbench selects a language model by name: the AI Model drop-down lists every configured model by its name, and the name you pick is sent with each request so that exact model configuration is used to answer it.

One of the configured models is designated the default LLM. The default is preselected in the AI Model drop-down when you start a new conversation, and it is also the model used as a fallback whenever a request does not name a specific model. This gives everyone a sensible model out of the box, while still letting you switch per conversation.

You can change which model is the default from the dashboard's AI configuration. As an alternative, the EvolveIT desktop setup utility also provides a Set Default LLM option: it lets an administrator choose the default model for the web application (and, separately, the default provider and model for the MCP Server), shows a preview of the changes, and applies them to the configuration. Either way, the Workbench then preselects the newly chosen default.

10. Step-by-step streaming of tool calls and results

Responses stream in live as they are generated rather than appearing only when finished. When step-by-step streaming is enabled in the AI configuration, the Workbench also surfaces each tool call as it happens: a collapsible card appears for every tool the AI invokes, showing its name and a Running status, then updating to Completed with the tool's result (or Stopped if you interrupt the request). Expand a card to inspect the exact arguments sent and the result returned.

This gives you full visibility into the AI's reasoning process (which tools it called, in what order, and what each returned) instead of just a final answer.