A new arXiv paper (2603.23533v4) takes aim at a common practice in retrieval-augmented generation: paying one LLM call per chunk during indexing. Pipelines often use these calls to enrich chunks with titles, summaries, keywords, and questions, or to add document-level context — an approach known as contextual retrieval (CR).

The authors frame the question directly: what does that extra call actually buy over free structure? They ask whether the added metadata and context meaningfully improve retrieval, or whether the structure already present in documents — headings, links, or other inherent signals — could do the job without the added cost.

The abstract stops before presenting any results, so the digest can only report the question, not the answer. The paper's contribution, as described, is in posing this cost-benefit comparison for a widely used but rarely scrutinized step in RAG pipelines.