Most of the retrieval pipeline complexity built into RAG systems may be wasted effort.
Researchers tested eight common retrieval enhancement methods — query expansion, hierarchical summarization, graph-based expansion, per-query routing, rank fusion, and corrective re-retrieval — against a shared backbone on both homogeneous and mixed-format document collections. To do this on realistic data, they built HetDocQA, a new benchmark that mixes code, markdown, tables, scientific PDFs, and prose in a single corpus, something existing benchmarks largely skip. The finding: a strong cross-encoder reranker already captures most of the quality gains, and once it is in place, nearly every add-on method stops paying off.
That matters because RAG pipelines in production almost always run on heterogeneous data, not clean Wikipedia prose — yet the research justifying these retrieval enhancements was done almost exclusively on homogeneous corpora. Teams adding graph expansion or rank fusion to their pipelines may be carrying complexity that benchmarks never actually validated for their use case. The one exception beyond query expansion is SSCC, a per-source calibrated corrector introduced in the paper, which sets separate acceptance thresholds per score source and shows reliable gains only on heterogeneous data.
The broader implication is familiar: more pipeline stages are not always better, and the research community's habit of benchmarking on Wikipedia has quietly understated how much a single strong component can dominate a system's output.