Improving Keyword Quality in a Production RAG Workflow

This engineering white paper presents a bounded, product-local evaluation of keyword extraction and filtering in a production RAG workflow. It compares a reference YAKE configuration, an improved deterministic YAKE-v2 pipeline, a full LLM filter, and a narrowly scoped LLM fragment filter. The results support YAKE-v2 as the production default while retaining the narrow LLM policy as an off-by-default, fail-open option.