Beyond Component Strength: Synergistic Integration and Adaptive Calibration in Multi-Agent RAG Systems

TL;DR

The paper investigates why adding powerful components to retrieval-augmented generation often fails to improve reliability when deployed in isolation. Through a controlled ablation on 50 queries, it shows that a hybrid retrieval, ensemble verification, and adaptive thresholding stack yields a 95% abstention reduction while keeping hallucinations in check, revealing emergent synergy. It also uncovers a labeling artefact that can misrepresent hallucination rates in ensemble setups and argues for standardized metrics and adaptive calibration. The findings advocate for integrated design and measurement frameworks as essential for deploying trustworthy multi-agent RAG systems in production settings.

Abstract

Building reliable retrieval-augmented generation (RAG) systems requires more than adding powerful components; it requires understanding how they interact. Using ablation studies on 50 queries (15 answerable, 10 edge cases, and 25 adversarial), we show that enhancements such as hybrid retrieval, ensemble verification, and adaptive thresholding provide almost no benefit when used in isolation, yet together achieve a 95% reduction in abstention (from 40% to 2%) without increasing hallucinations. We also identify a measurement challenge: different verification strategies can behave safely but assign inconsistent labels (for example, "abstained" versus "unsupported"), creating apparent hallucination rates that are actually artifacts of labeling. Our results show that synergistic integration matters more than the strength of any single component, that standardized metrics and labels are essential for correctly interpreting performance, and that adaptive calibration is needed to prevent overconfident over-answering even when retrieval quality is high.

Figures (4)

• Figure 1: Hallucination (left) and abstention (right) rates across configurations for the 25 benign plus edge‑case queries. Ensemble‑only appears to hallucinate 40% of the time due to a labelling artefact, while all other configurations exhibit 0% hallucinations. Baseline, hybrid‑only and adaptive‑only configurations abstain on 40% of queries, whereas the full‑stack abstains on just 2%.
• Figure 2: Confidence‑tier distributions across all configurations. Ensemble‑only (top right) shows 100% high‑confidence predictions, explaining its overconfidence. The full stack (bottom right) exhibits more balanced distributions with high confidence for most queries and medium confidence for a few edge and adversarial cases.
• Figure 3: Average latency by configuration. The full stack (with hallucination metrics) is significantly slower due to additional verification steps, whereas the other configurations remain under 10 seconds.
• Figure 4: Performance versus latency trade‑off. Performance is defined as $100 - 100 \times \text{hallucination rate} - 50 \times \text{abstention rate}$. Ensemble‑only falls into the "slow and poor" quadrant due to mislabelled hallucinations, while baseline and hybrid‑only are fast but still poor due to high abstention. Adaptive‑only matches the baseline performance but is slightly slower.