Prefix every chunk with its own header chain: +23.8% MRR@5 for zero extra LLM calls, plus a measurement trap that invalidates common chunking evals
A three-stage, chunk-side-only pipeline — header-split, semantic merge, then title-chain prefixing that reuses the document's existing header hierarchy instead of a generated summary — lifts MRR@5 from 0.374 to 0.463 (+23.8%) on a 1,600-query stratified evaluation over a production Markdown knowledge base, and from 0.828 to 0.925 on the answerable subset, at zero additional inference cost. The paper also reports what failed, including the methodological finding: in a same-pool prefix on/off ablation, dual-annotator agreement collapsed from Cohen's kappa 0.45 to 0.04 under identical prompts, because stripping the title-chain context also strips the signal annotators need to agree on relevance. If you have ever A/B'd a text transform in a RAG pipeline by re-scoring the same pool with and without it, that result says your comparison may have been measuring annotator confusion.
↳ Follow the thread