Guardrails

Currently in progress...

Using a 'support' chatbot type app for financial or thrapeutics, how do we evaluate the quality of the guardrails for PII, safety, ethics and tone of our Agents with particular emphasis on the overall workflow of Evals as a whole.

It is linear as all nodes must be executed and at any point the APP can exit with a failure. Branching is an option too.

Additional nodes can be added to make the workflow more complex.

Here we can see the second node that has PII already done giving its decision is_self_destructive: