rebuff
rebuff copied to clipboard
protectai
LLM Prompt Injection Detector
It is possible to evade the model check (make it always output score 0.0) by appending a special suffix to a prompt. The prompt injections then are not detected by...
To support more online/local models and vector stores, it will be more convenient to pass into langchain components, so that we don't need to handle initiation of various models and...
in v0.1.1, `pinecone_environment` was added as an arg for rebuff, which was later passed to the `pinecone.init` function. When using the latest pinecone version, this breaks with teh following error:...
As encountered in https://github.com/protectai/rebuff/issues/68, it's possible for the language model to not return a numerical value. With the python sdk, this causes an exception, but with the js sdk, this...
I realized that in #90, I left the `javascript-sdk` in a state where it doesn't compile. My reasoning for not including `javascript-sdk/src/api.ts` in that PR was because the server API...
We should have API docs for Python and TypeScript SDKs (After #52), a getting started guide that walks users through setting up their infrastructure (After #10) , examples in how...
According to the script `detect_pi_vectorbase.py` in `python-sdk/rebuff/`, the below code snippet assumes higher similarity score implies prompt injection, whereas it seems it should be the opposite i.e. a lower similarity...
This PR updates the server code to match the changes made in https://github.com/protectai/rebuff/pull/90. This PR includes the commit(s) from that PR along with a commit from https://github.com/protectai/rebuff/pull/66 which was needed...