Code execution
Stevin can write and run a short Python program to compute over data it already has: total, join, rank, check arithmetic, or produce a file.
What it is for#
Language models are unreliable at arithmetic over many rows. For that kind of work Stevin writes a short Python program and runs it, then reports the result. Typical uses:
- Total, average or rank the rows of a table.
- Join two lists on a shared field.
- Check the arithmetic in a quote or a report.
- Produce a file: a spreadsheet, a CSV, a chart.
You do not have to ask for code. Ask for the result ("total the hours per project and give me a spreadsheet") and Stevin decides when a program is the right tool.
What the program can and cannot reach#
The program runs in a sandbox, separate from your systems:
- It works on data Stevin already has in the conversation or has fetched through your connected tools.
- It has no network access, so it cannot call outside services or send data anywhere by itself.
- Files it produces come back through Stevin, are attached to the reply and are kept in the library.
Installing packages#
When a program needs a library the sandbox does not have, Stevin can install a Python package into the sandbox.
- Installing a package asks for approval, like any other action with an effect.
- An admin can list package names that install without approval (for example
pandas,openpyxl), on the Install packages integration. - Who may have packages installed, and who may change the list, are two separate permissions.
Turn it on or off#
Code execution and Install packages are listed under Customize → Integrations in the console. An admin can connect or disconnect each one. See Integrations.