From d39d7f1f1f848049522a265cbf1d02e066c9b6fa Mon Sep 17 00:00:00 2001 From: Adam Wilson Date: Mon, 28 Jul 2025 13:29:21 -0600 Subject: [PATCH] don't send back alternate after guidelines --- .../guardrails/reflexion_security_guidelines_service.py | 2 +- .../services/nlp/text_generation_completion_service.py | 4 ---- 2 files changed, 1 insertion(+), 5 deletions(-) diff --git a/src/text_generation/services/guardrails/reflexion_security_guidelines_service.py b/src/text_generation/services/guardrails/reflexion_security_guidelines_service.py index cb1102858..c091d75e5 100644 --- a/src/text_generation/services/guardrails/reflexion_security_guidelines_service.py +++ b/src/text_generation/services/guardrails/reflexion_security_guidelines_service.py @@ -142,7 +142,7 @@ class ReflexionSecurityGuardrailsService( result.alternate_result = None return result - # provide the finalized alternate (refuse to answer) + # last resort; provide the finalized alternate (refuse to answer) print(f'Guardrails-based completion was malicious. Score: {result.guardrails_result.semantic_similarity_result.max}') result.alternate_result = AlternateCompletionResult( alterate_completion_text = self.constants.ALT_COMPLETION_TEXT diff --git a/src/text_generation/services/nlp/text_generation_completion_service.py b/src/text_generation/services/nlp/text_generation_completion_service.py index bae834feb..1fdb1f3c1 100644 --- a/src/text_generation/services/nlp/text_generation_completion_service.py +++ b/src/text_generation/services/nlp/text_generation_completion_service.py @@ -128,11 +128,7 @@ class TextGenerationCompletionService( print(f'Guidelines-based completion was NOT malicious. Score: {completion_result.guidelines_result.semantic_similarity_result.max}') return completion_result - # provide the finalized alternate (refuse to answer) print(f'Guidelines-based completion was malicious. Score: {completion_result.guidelines_result.semantic_similarity_result.max}') - completion_result.alternate_result = AlternateCompletionResult( - alterate_completion_text = self.constants.ALT_COMPLETION_TEXT - ) completion_result.finalize_completion_text() return completion_result