We detect contradiction between query and response using embedding cosine similarity — no extra sampling, no NLI models.
Input-Output Embedding Contradiction Detection
The Problem
Contradiction detection in LLM outputs typically requires either sampling multiple responses (SelfCheckGPT) or running a separate NLI model (DeBERTa, BART). Both approaches add latency and memory that a local-first system cannot afford.
