RAILS method raises LLM clustering scores and reaches production deployment

A preprint introduces RAILS, a system that clusters documents with a large language model while retrieving labels from a growing pool and processing documents in concurrent batches. Across six public benchmarks, the authors report that it raised average accuracy from 51.2% to 59.3%, normalized mutual information from 67.2% to 74.8%, and adjusted Rand index from 45.4% to 54.7%. The researchers also report replacing HDBSCAN in a software provider's ticket-topic discovery pipeline. The evidence comes from the paper's authors, and the abstract does not identify the provider or supply independent production measurements.