We are building a patient support bot and must ensure no PII is leaked. How reliable is the Guardrails AI PII filter for medical contexts? I’m concerned about it accidentally redacting medical terminology that might look like names or addresses, which would ruin the helpfulness of the bot's advice.
3 answers
In a medical setting, you should use the Presidio-based validator within Guardrails AI. It allows for high-granularity control over what entities are flagged. To avoid "over-redaction" of medical terms, you can pass a 'deny_list' or an 'allow_list' to the validator. For example, you can explicitly tell the guard to ignore common drug names or anatomical terms if the NER model misidentifies them as names. It is also a good practice to use the 'Fix' strategy with a placeholder like [REDACTED_NAME] rather than just deleting the text, so the doctor or patient can still follow the logic of the response while the sensitive data is obscured.
Are you planning to run these guards on the input (user prompt) or just the output (LLM response)? HIPAA usually requires both.
You can also use custom regex patterns for specific medical ID formats that standard PII filters might miss. It adds that extra layer of safety.
Definitely, Harrison. Custom regex for internal hospital IDs has been a lifesaver for our specific compliance requirements.
We are planning to implement them on both ends. Protecting the output is the final safety net, but redacting the prompt before it even hits the LLM is our primary goal to prevent sensitive data from being stored in the model provider's logs. We want to be 100% compliant before we even think about moving out of the beta testing phase.