Answers are grounded in this paper and cite their sources.
Retrieval-Augmented Generation (RAG) pipelines repeatedly make small judgments, such as whether a passage is relevant, whether the retrieved context is sufficient and whether an answer is grounded. These are usually delegated to the generating large language model (LLM), at notable latency and cost, or to score thresholds without a probabilistic interpretation. We study nano-jev, a 33.4\,M-parameter cross-encoder that treats relevance (3-way), sufficiency and groundedness (binary) as typed, closed-set classification problems and returns temperature-calibrated probabilities. Each (question, option, context) triple is scored by a single-logit head, which makes predictions invariant to option order. On held-out halves of the HotpotQA, SQuAD~2.0 and MultiNLI validation sets, nano-jev reaches mean accuracies of 0.812 (relevance), 0.843 (sufficiency) and 0.841 (groundedness) over five seeds, comparable to a 1-pass classifier trained on the same data, which is 2.5× faster. Dedicated checkers such as MiniCheck remain stronger on groundedness (0.866). nano-jev runs 55-93× faster than a prompted 4\,B-parameter LLM. On unseen datasets (MuSiQue, VitaminC), accuracy drops by 15-20 points and ECE rises from about 0.03 to about 0.17, so target-domain re-calibration is advisable. A confidence cascade that escalates uncertain decisions to the LLM improves in-distribution groundedness (0.868 vs. 0.841, $p=0.02$) with about a quarter of decisions escalated, but trades accuracy for latency out of distribution. In a small end-to-end test, nano-jev relevance filtering improves HotpotQA answer F1 by 2.1 points at about 39ms per question. Code and weights are released as an open-source package.
No takes yet. Share an insight, caveat, or question.
Divakar et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: