bug: Jailbreak detection imports unnecessary dependencies when using NIM

[jeffreyscarpenter]

Did you check docs and existing issues?

• I have read all the NeMo-Guardrails docs
• I have updated the package to the latest version before submitting this issue
• (optional) I have used the develop branch
• I have searched the existing issues of NeMo-Guardrails

Python version (python --version)

Python 3.12.0

Operating system/version

MacOS 15.5

NeMo-Guardrails version (if you must use a specific version and not the latest

0.14.0

Describe the bug

The jailbreak detection module imports heavy dependencies (torch, transformers, sklearn) at module level, even when using the NIM-based approach. This means users need to install these dependencies even if they're not using the local model-based jailbreak detection.

Steps To Reproduce

1. Install NeMo-Guardrails without torch, transformers, or sklearn
2. Configure jailbreak detection to use NIM-based approach:

rails:
  config:
    jailbreak_detection:
      nim_url: "your-nim-url"
      nim_port: 8000

3. Try to use the jailbreak detection feature
4. Observe that the code fails to import required dependencies

Expected Behavior

When using the NIM-based approach for jailbreak detection, the code should not require torch, transformers, or sklearn to be installed. These dependencies should only be required when using the local model-based approach.

Actual Behavior

The code fails with import errors for torch, transformers, or sklearn even when using the NIM-based approach, because these dependencies are imported at module level rather than being lazy-loaded when needed.