Classification is a RAG problem: A case study on hate speech detection (Hugging Face)

Adaptable classification algorithms that react to changing regulations without frequent retraining are necessary for effective content control. In this work, a Retrieval-Augmented Generation (RAG) method is presented, redefining classification as policy-based evaluation rather than fixed category prediction. This moves the focus of hate speech identification to determining whether content contravenes particular policy rules. Three main advantages of the suggested Contextual Policy Engine (CPE), an agentic RAG system, are smooth policy updates, inherent explainability via recovered policy segments, and competitive classification accuracy. RAG’s promise for adaptable and transparent content moderation is demonstrated by experimental results that show that CPE permits fine-grained control over identity group safeguards while preserving overall performance.

https://huggingface.co/papers/2508.06204