Rspamd: New Frontiers in Spam Detection with Embeddings and Neural Networks

This talk covers recent Rspamd developments with a focus on leveraging modern ML techniques for email filtering. We'll explore HTML fuzzy hashing — a structural fingerprinting approach for detecting phishing templates — and multi-class Bayesian classification supporting up to 20 mail categories in a single Redis lookup. The main part of the talk will dive into the new provider-based neural network architecture that enables combining multiple feature sources, and the integration of LLM embeddings as a powerful signal for spam classification. We'll discuss practical deployment considerations, embedding model selection, and how neural network outputs can be shared across Rspamd instances. The talk will also touch on our experience using LLM tools during development — what worked, what didn't, and where human judgment remains essential.