🤔 What is the technical debt you think should be addressed?
Commit 05f71b3f3 (feat(responses): enable default memory vector store) changed the sentence-transformers embedding provider configuration in the starter distribution from trust_remote_code: false to trust_remote_code: true. This affects:
- src/ogx/distributions/starter/config.yaml
- src/ogx/distributions/starter/run-with-postgres-store.yaml
- src/ogx/distributions/starter/starter.py
The config class itself documents this flag as defaulting to false for security reasons:
▎ "Whether to trust and execute remote code from model repositories. Set to True for models that require custom code (e.g., nomic-ai/nomic-embed-text-v1.5). Defaults to False for security."
With trust_remote_code: true, the sentence-transformers provider will execute arbitrary Python code bundled with any model loaded from the HuggingFace Hub. A compromised or malicious model repository could use this to achieve remote code execution on the OGX server.
Expected behavior: The starter distribution should default to trust_remote_code: false, matching the field's own default and the previous behavior.
Suggested fix: Revert the three files above to trust_remote_code: false. If a specific model needs it, document that in the model-specific setup instructions instead.
💡 What is the benefit of addressing this technical debt?
This change appears unrelated to the memory vector store feature that was the focus of the PR. If a specific embedding model requires trust_remote_code, the recommendation should be scoped to that model's documentation rather than enabled by default in the starter distribution, which is the template most new deployments start from.
Other thoughts
No response
🤔 What is the technical debt you think should be addressed?
Commit
05f71b3f3(feat(responses): enable default memory vector store) changed the sentence-transformers embedding provider configuration in the starter distribution fromtrust_remote_code: falsetotrust_remote_code: true. This affects:The config class itself documents this flag as defaulting to false for security reasons:
▎ "Whether to trust and execute remote code from model repositories. Set to True for models that require custom code (e.g., nomic-ai/nomic-embed-text-v1.5). Defaults to False for security."
With
trust_remote_code: true, the sentence-transformers provider will execute arbitrary Python code bundled with any model loaded from the HuggingFace Hub. A compromised or malicious model repository could use this to achieve remote code execution on the OGX server.Expected behavior: The starter distribution should default to trust_remote_code: false, matching the field's own default and the previous behavior.
Suggested fix: Revert the three files above to
trust_remote_code: false. If a specific model needs it, document that in the model-specific setup instructions instead.💡 What is the benefit of addressing this technical debt?
This change appears unrelated to the memory vector store feature that was the focus of the PR. If a specific embedding model requires
trust_remote_code, the recommendation should be scoped to that model's documentation rather than enabled by default in the starter distribution, which is the template most new deployments start from.Other thoughts
No response