Perhaps an idea for more enterprise features in LibreChat

[dirkpetersen]

https://robert-mcdermott.medium.com/adding-a-safety-layer-in-generative-ai-systems-ed0aaa96a6c1

The article is about

• implementing a safety layer in generative AI systems to prevent toxic behavior and disclosure of sensitive information.
• It proposes using lightweight BERT models as middleware between user interfaces and backend language models to detect, filter, block, or de-identify content.
• demonstrates using the 'detoxify' package to detect harmful content and block toxic language in a chat application.
• explores using BERT models to detect and de-identify personally identifiable information (PII) and protected health information (PHI) in medical notes.
• implemented both toxicity and PHI-blocking features in their custom ChatGPT-like platform, finding the toxicity filter effective but the PHI detector more suitable for specific use cases rather than general-purpose chat applications due to false positives.