Retriever-Agnosticism of Distribution-Shape Features

Determine whether the distribution-shape features used by the GeneralQPP retrieval-sufficiency predictor remain effective across different retrievers, beyond the single retriever evaluated in the study.

Background

GeneralQPP predicts RAG retrieval sufficiency using non-lexical features derived primarily from the shape of the retrieval-score distribution. The experiments evaluate these features with one retriever, Qwen3-VL-Embedding-8B, across the ViDoRe and SÚJB datasets. Although the method performs strongly in the reported settings, its effectiveness may depend on the score geometry produced by that particular retriever.

The paper explicitly identifies retriever-agnosticism as unresolved because no direct evaluation across multiple retrievers was conducted. Establishing whether the features generalize across retrievers is important for determining whether GeneralQPP is a broadly applicable QPP method or is specific to the tested retrieval model.

References

With a single retriever and a single local judge in our setup, retriever-agnosticism of the distribution-shape features is an empirical prediction we have not tested directly.

— Your Retriever Already Knows: Distribution-Shape QPP for RAG Retrieval Sufficiency  (2609.11646 - Veselý et al., 10 Sep 2026) in Section Discussion, paragraph “Limitations”