Analyses, spreadsheets and code

“Turn these numbers into an Excel analysis for me” is one of the tasks your team is most likely to need an AI for. This page describes what happens along the way, which part works without additional technology and which part requires a backend.

Code execution and artifacts come from LibreChat. What we add is the selection and preparation of the execution backend.

What you get out of it

You write: “From this list please make an Excel file with one analysis per month.” The colleague calculates, creates the file and puts it into the chat as a download. You click it and you have the file.

This applies to Excel, CSV, Word and PDF alike, and also to analyses of an attached table. Only one thing matters: say what should be in the file, not how it should technically come into being.

If you want to change a file that has already been created, say so in the same chat. The colleague loads the existing file and adapts it instead of rebuilding it. That way what you agreed on earlier is preserved.

The path to the code stays in the background

Scripts, commands and internal paths are not shown. You get the file and a short summary of what is in it. That is deliberate: the target group is professionals, not developers.

Artifacts: results directly in the chat

Artifacts are something different from code execution. With them the colleague presents content directly in a side area of the interface instead of emitting it as a code block:

Type	Example
Mermaid	Flowchart, org chart, timeline
SVG	simple graphics and schematics
HTML	a designed page or table to look at
React	a small, operable interface, for example a calculator
Markdown	formatted documents

Artifacts need no execution backend. They are active on the shipped model cards and work immediately after installation. So anyone who only wants a diagram or an overview does not need code execution.

The preview needs HTTPS

The interactive preview only runs in a secure context. Without a valid certificate it stays empty or reports a connection error. That is not a fault of the backend but a browser rule. The clean route is a real domain name with a valid certificate.

Code execution is off in the shipped state

Out of the box, code execution is not enabled. The reason is the same as with file search: execution needs a separate, sealed off backend that is not part of the base installation. Without a backend the button would be a dead button and every attempt would end in an error.

So you decide deliberately whether to offer this feature.

For admins

Code execution is the only place where generated code runs on your machine. Accordingly it belongs in an isolated environment of its own, operated separately from the rest of the installation and never shared between instances.

Three points decide when picking a backend:

• Isolation. Executed code must reach neither the data of the instance nor the rest of the network.
• Reachability. The backend is addressed internally only. Nothing is needed from outside for that; diagnostic access belongs on the loopback address or in an SSH tunnel.
• Outbound traffic. Which destinations executed code may reach on the internet is defined at host or network level. Plan this allowlist as part of the setup and record the permitted destinations in your operations handbook.

If code execution is not needed at all, leave the backend out: it is switched off out of the box.

Enabling it afterwards

1. Set the permission. The permission “Code-Interpreter” (code interpreter) is in the role editor under “Werkzeuge” (tools). If you want to give execution to individual roles only, maintain it exclusively there.

2. Activate it on the model cards so that the model uses the tool by itself, without anyone clicking.

3. Check that a generated file really appears as a download in a fresh chat.