Risk-Controlled Adaptive Multi-Hop RAG with Evidence-Chain Retrieval, Word-Level Hallucination Detection, Attribution, and Calibrated Abstention

Megan Walters (University of California, Santa Cruz)
Qing Song (Vanderbilt University)
Austin Jiang (University of California, Riverside)

Journal of Information Systems and Business Technology
Vol 2 No 4 (2026): Journal of Information Systems and Business Technology

Reliable RAG fails when retrieval omits evidence, generation is unsupported, or confidence is misplaced. We evaluated all 2,556 MultiHop-RAG questions and 2,700 official RAGTruth test responses (2,675 good-quality in primary analysis). On chain-group-disjoint MultiHop-RAG, adaptive retrieval achieved 0.8810 Recall@k and 0.6925 complete-chain rate with 10.85 documents per answerable query; the nearest lower-cost comparator reached 0.8463 and 0.6273. Answer token F1 was 0.5562 and strict URL attribution 0.7753. On RAGTruth, the calibrated detector reached 0.3254 word F1, 0.0986 overlap span F1, 0.00582 ECE, and 0.03550 Brier score. Selective release cut risk from 0.3525 at full coverage to 0.2140 at 80%. Evidence, grounding, provenance, and release risk require separate coordinated measurement.