Code interpreter with Jupyter Notebook does not work if generated code is executed again via "Run" Button

[InventoCasa]

Bug Report

Installation Method

Docker

Environment

• Open WebUI Version: 0.5.11

• Ollama (if applicable): -

• Operating System: Ubuntu 22.04

• Browser (if applicable): Chrome

Confirmation:

• I have read and followed all the instructions provided in the README.md.
• I am on the latest version of both Open WebUI and Ollama.
• I have included the browser console logs.
• I have included the Docker container logs.
• I have provided the exact steps to reproduce the bug in the "Steps to Reproduce" section below.

Expected Behavior:

When clicking on "Run" in the code block of the answer of the LLM, the code should be executed again with Jupyter notebook

Actual Behavior:

It seems like the code is executed through the standard Pyoide Code interpreter.
Although I am not sure, because in my case the pyoide interpreter does not work at all since v0.5 - but this leads to "Execution timeout exceeded" when running the generated code again, so I assume the Pyoide interpreter is chosen instead of Jupyter. Also, I see no call in the jupyter logs when executing the code via "Run" again. On the initial run which is done right when the answer is generated by the LLM, is see in the jupyter logs that it is working

Description

Bug Summary:
Code interpreter with Jupyter Notebook does not work if generated code is executed again via "Run" Button

Reproduction Details

Steps to Reproduce:

1. Setup Jupyter Code Interpreter
2. Select Code Interpreter in chat and prompt e.g. "Create a script which sums up 3 random numbers"
3. Observe that the answer correctly calls jupyter notebook and returns the execution output of the generated code
4. Click on "Run" again in the code block, and observe jupyter logs: Nothing will be logged. As mentioned in my case, Pyoide does not work since v0.5 (despite setting up websocket modifications correctly in nginx), that's why I noticed this bug.

[dheshanm]

I notice similar behaviour:

Steps to Reproduce:

1. Setup Jupyter Code Interpreter
2. Select Code Interpreter in chat and prompt: "Print todays’ date in different colrs, highlighting month, day and year in different colors using python"
3. Code gets generated (see below for generated code and output)

Generated Code for reference:

from datetime import datetime
from termcolor import colored

# Get today's date
today = datetime.now()

# Extract month, day, and year
month = today.strftime('%B')  # Full month name
day = today.day
year = today.year

# Print the date with different colors
colored_month = colored(month, 'red')
colored_day = colored(day, 'green')
colored_year = colored(year, 'blue')

print(f"Today's date is: {colored_month} {colored_day}, {colored_year}")

STDOUT/STDERR

PythonError: Traceback (most recent call last):
File "/lib/python312.zip/_pyodide/_base.py", line 597, in eval_code_async
await CodeRunner(
File "/lib/python312.zip/_pyodide/_base.py", line 411, in run_async
coroutine = eval(self.code, globals, locals)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "<exec>", line 2, in <module>
ModuleNotFoundError: No module named 'termcolor'
The module 'termcolor' is included in the Pyodide distribution, but it is not installed.
You can install it by calling:
await micropip.install("termcolor") in Python, or
await pyodide.loadPackage("termcolor") in JavaScript
See https://pyodide.org/en/stable/usage/loading-packages.html for more details.

Expected Behavior:

The generated code snippet should have been executed in the Jupter notebook environment, which had these dependencies installed.

It would also be helpful if each execution also had some indication on which interpreter (pydiode / jupyter) was used. I was expecting jupyter was used on my other snippets, but it seems all code interpreter calls have been directed to pydiode, despite having jupyter server configured.

[tjbck]

Intended behaviour here for security reasons.

[InventoCasa]

Intended behaviour here for security reasons.

Can you explain that further, why is this for security reasons? This just doesn't make sense to me..

Also isn't this terrible from a User perspective? Being able to execute the code just once with jupyter, then showing the "Run" button and then it get's executed in pyoide which is missing the libraries installed in Jupyter?

[alesha-pro]

@tjbck you should fix this, I suppose, because its literally doesnt make sense

[FlorinAndrei]

@tjbck The Jupyter feature literally does not work. It uses pyodide no matter what. Please fix, or provide an explanation.

[dkapur17]

@FlorinAndrei, the Jupyter code interpreter does work if you have the model generate and execute code on its own. You can try the following prompt with a model that is allowed to use the code interpreter:

"Can you write and execute python code printing the current executable?"

Within the Analysis collapsible, you'll see that its using the Jupyter interpreter (same as if you were to run which python in the Jupyter server)

However, if you were to click on Run and re-run the code block, it would instead use Pyodide:

Again, you can validate this by installing some package on your Jupyter server and then asking the model to execute code importing that package.

[FlorinAndrei]

@dkapur17 Thank you for the explanation.

The problem is - this makes the feature much less useful. Could you guys put a bunch of "ARE YOU SURE YOU WANT TO DO THIS DANGEROUS THING" checkboxes in the settings, which, when selected, would give users like me all the dynamite we need to blow ourselves up with?

[alesha-pro]

@dkapur17 Thank you for the explanation.

The problem is - this makes the feature much less useful. Could you guys put a bunch of "ARE YOU SURE YOU WANT TO DO THIS DANGEROUS THING" checkboxes in the settings, which, when selected, would give users like me all the dynamite we need to blow ourselves up with?

Yes. Add the warning or something else.

I am creating a dedicated container with jupyter just for this purpose and would like to be able to use it to its full potential. Please. It would be a good option to be able to restart the code in jupyter.
Re-open this issue.

[dkapur17]

Haha, @FlorinAndrei @anonymousmaharaj I'm in the same boat as you guys. Came across this issue because I too was looking for the solution. I'm not a contributor 😅

[alesha-pro]

Haha, @FlorinAndrei @anonymousmaharaj I'm in the same boat as you guys. Came across this issue because I too was looking for the solution. I'm not a contributor 😅

While we r were talking, they made https://github.com/open-webui/open-webui/releases/tag/v0.5.13.

[dheshanm]

The run button still (v0.5.13) uses the Pyodide interpreter:

Using dkapur17's prompt, on pressing the run button from the snippet:

[alesha-pro]

The run button still (v0.5.13) uses the Pyodide interpreter:

Using dkapur17's prompt, on pressing the run button from the snippet:

Check settings for executor. There is 2 options now to enable jupyter

[jamorris]

This wasn't obvious to me but since I'm running both Open WebUI and Jupyter Notebook in Docker I had to use the Jupyter container's internal Docker network address (172.17.x.x) when configuring Open WebUI's Code Execution settings. That's in the container's Inspect tab in Docker. Once I set that up it works beautifully.