Open msintaha opened 2 months ago
We're looking at this, and we think that we can get rebuff to work in one of the following ways
As far as the idea of using a diff prompt injection model from huggingface goes, I think this is something we should pursue regardless. If you would like to help create that validator, I can show you how and you can share it on the hub. Otherwise, we might take a bit to get to it
We're looking at this, and we think that we can get rebuff to work in one of the following ways
- add the github release to the pyproject
- add a pip installation from github to the post_install script
As far as the idea of using a diff prompt injection model from huggingface goes, I think this is something we should pursue regardless. If you would like to help create that validator, I can show you how and you can share it on the hub. Otherwise, we might take a bit to get to it
@zsimjee thanks for getting back on this. For integrating the huggingface models, if there's some guideline on getting it integrated with a validator, I can take a look when I have some time, but for the rebuff workaround, I hope that's something coming up soon! Can you share timelines for this change?
@msintaha sweet yes the rebuff validator fix is merged, but I haven't validated it yet. Should be able to try again, otherwise I'll validate it later tonight.
For huggingface models, you can follow the pattern here -
validator code - https://github.com/guardrails-ai/sensitive_topics/blob/67adbe4f47716b61a61486ad7fc7f87766f86d89/validator/main.py#L111
post-install script - https://github.com/guardrails-ai/sensitive_topics/blob/main/validator/post-install.py
validator - https://hub.guardrailsai.com/validator/guardrails/sensitive_topics
@zsimjee I just tested by pulling the DetectPromptInjection
validator again, it seems the error for the positional argument being required (pinecone_environment
) still exists
Describe the bug The DetectPromptInjection library is currently unusable because the underlying library (Rebuff) has not released the change for their API initialization. It seems you have reached out to them on their github repo, but I was wondering if the dependency on Rebuff can be changed to an alternative option as they dont seem to be responsive. Perhaps a pre-trained huggingface model?
To Reproduce Adding
pinecone_api_key
following the steps defined in DetectPromptInjection validator throwsRebuffSdk.__init__() missing 1 required positional argument: 'pinecone_environment'
errorExpected behavior Works without defining the
pinecone_environment
or using a different underlying library to detect prompt injectionsLibrary version: Latest