[Rasa Pro Bug] Persistent OSError: [WinError 123] on Windows During Model Loading

Getting Started with Rasa
1

Hello Rasa Community and Support Team,

I am encountering a persistent and blocking OSError: [WinError 123] when trying to run a trained model on Windows using Rasa Pro. The training process completes successfully, but I cannot interact with the model using rasa shell or rasa run.

I have followed a very detailed troubleshooting process and believe this may be a bug in how model archives are handled on Windows in this version.

Environment Details

  • **Rasa Pro Version : 3.12.17
  • **Minimum Compatible Version: 3.11.0rc1
  • **Rasa SDK Version : 3.12.0
  • **Python Version : 3.10.18
  • **Operating System : Windows-10-10.0.26100-SP0
  • Installation Method: uv pip install rasa-pro.

Problem Description

  1. rasa train runs and completes successfully, creating a model in the models/ directory.
  2. rasa run actions starts the action server successfully, registering all custom actions.
  3. When I run rasa shell, it immediately fails while trying to load the model.

Error Log from rasa shell:

ERROR rasa.main - [error] OSError: [WinError 123] The filename, directory name, or volume label syntax is incorrect: ‘\\?\C:\Users\v4var\AppData\Local\Temp\tmphn0ukstr\components/train_TEDPolicy1’ event_key=cli.exception.general_exception

… (full traceback) …`

The key error is OSError: [WinError 123] which seems to be caused by an invalid file path mixing Windows (\) and Unix (/) separators during the model extraction process into a temporary directory.

Troubleshooting Steps Already Taken

Based on extensive debugging, we have tried the following workarounds without success:

  1. Created a Clean Environment: The current rasa-pro-env was created from scratch with conda create --name rasa-pro-env python=3.10 to ensure a compatible Python version.
  2. Confirmed rasa-pro Installation: uv pip install rasa-pro was run successfully in the new environment.
  3. Manual Model Unpacking: To bypass the faulty extraction logic, we manually unpacked the .tar.gz model into a folder named unpacked_model.
  4. Verified Unpacked Structure: We confirmed the unpacked_model folder correctly contains the Rasa 3.x model structure: a components/ folder and a metadata.json file.
  5. Attempted to Load Unpacked Model: We then tried to run the server by pointing directly to this unpacked directory using rasa run --model unpacked_model.
  6. Final Error: This command still fails, but with a different error: No model found.. This happens even after verifying the directory contents and structure are correct with ls unpacked_model immediately before running the command. Using an absolute path in the --model flag also results in No model found.

Core Question

Given that rasa train works, but loading the model fails with a file path error (WinError 123) and the standard workaround (--model <directory>) also fails (No model found), this seems to be a blocking bug for local development on Windows.

Is this a known issue? What is the recommended procedure to load and interact with a trained Rasa Pro model locally on a Windows machine in light of these errors?

Thank you for any help or guidance you can provide.

2

Hi @Ruchi30 thank you for your detailed message. The Rasa team is investigating this issue.

Have you tried using any earlier versions of Rasa?

Also, to interact with the assistant, it’s best to use rasa inspect instead of rasa shell. Could you try this as well?

3

Hi Ashurkina… Thanks for your reply. I am getting same error with rasa inspect.