Root Cause Analysis (RCA)

To enhance transparency in our guardrail detection models, we’ve introduced a new feature: Root Cause Analysis (RCA). This feature provides a concise explanation behind the decisions made by our guardrailing models.

You can find the RCA feature in our dashboard, where it is presented as follows:

Sample of our RCA feature

Currently, RCA is enabled for the following metrics:

Metrics	RCA Enabled
ANSWER_RELEVANCE	✅
FACTUAL_CONSISTENCY	✅
CONCEPTUAL_SIMILARITY	✅
RESPONSE_TONE	✅
PROMPT_INJECTION	✅
DATA_LEAKAGE	✅
READABILITY	✅
INSECURE_OUTPUT	✅
INVISIBLE_TEXT	✅
BIAS	✅
COHERENCE	✅
DIVERSITY	✅
CREATIVITY	✅
GRAMMATICAL CORRECTNESS	✅
CLARITY	✅
NARRATIVE CONTINUITY	✅
TOXICITY	✅
BLEU_SCORE	✅
COSINE_SIMILARITY_SCORE	✅
COMPRESSION_SCORE	✅
FUZZY_SCORE	✅
METEOR_SCORE	✅
ROUGE_SCORE	✅