Guardrails: replacement text per word, or rewrite the sentence instead of swapping the word
T
Thomas Mueller
A "Blocked words" guardrail has one replacement text for all its words, and this breaks sentences. With "." as the replacement, the guest read "ob unser Haus eine . ist". Without a replacement text the guest sees "****".
Request: a replacement text per blocked word, or an option for the guardrail to ask the model to rewrite the sentence instead of swapping the word. Please also document what "Block the message" does exactly: no reply, a fixed reply or a new attempt.
Impact: guardrails that protect the guest without producing broken sentences.
Log In